BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//6.5.2//EN
TZID:Asia/Kolkata
X-WR-TIMEZONE:Asia/Kolkata
BEGIN:VEVENT
UID:72@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240830T160000
DTEND;TZID=Asia/Kolkata:20240830T170000
DTSTAMP:20240809T094946Z
URL:https://cds.iisc.ac.in/events/seminar-cds-202-0400-august-the-circulan
t-decomposition-of-a-matrix-and-fast-multiplication-of-large-matrices/
SUMMARY:{Seminar} @ CDS: #202: 04:00 August: "The circulant decomposition o
f a matrix and fast multiplication of large matrices."
DESCRIPTION:Department of Computational and Data Sciences\nDepartment Semin
ar\n\n\n\nSpeaker : Murugesan Venkatapathi\,Department of Computational &a
mp\; Data Sciences.\nTitle : The circulant decomposition of a matrix and f
ast multiplication of large matrices.\nDate &\; Time : August 30\, 2024
\, 4:00 PM\nVenue : # 202\, CDS Class room\n\n\n\nABSTRACT\nThe well-known
singular-value decomposition of a matrix\, and the eigenvalue decompositi
on of a non-defective square matrix\, have become indispensable in enginee
ring and sciences. Here\, the matrix is a weighted sum of rank-one matrice
s with the corresponding singular-value or the eigenvalue as the weight. T
ypically\, a decomposition is useful if a few significant components in th
e sum contain most of the required information in the matrix.\n\nAnother d
ecomposition of a n x n matrix was proposed here at the Institute\, where
the matrix is a sum of 'n' circulant matrices (with fixed periodic relaxat
ions on the unit circle). This decomposition is orthogonal with respect to
a certain inner product of matrices and allows a simple projection for th
e circulant matrices. Alternately\, a more efficient evaluation of all cir
culant components is also possible in O(n^2 . logn) operations exploiting
the Fast-Fourier-Transform (FFT). Note that a circulant matrix has at most
'n' unique entries cyclically permuting both in its rows and columns. Rel
atively few dominant circulant components may be sufficient in approximati
ng a dense matrix when it has some periodicity in its entries. For such ma
trices\, this decomposition was used in approximating eigenvalues and spar
se similarity transformations\, and to precondition linear solvers.\n\nMor
e generally\, the circulant decomposition was recently used to demonstrate
iterative multiplication of two matrices in O(n^2 . logn^2) i.e. Õ(n^2)
arithmetic operations\, with well-bounded relative errors less than 1% unl
ike other restricted methods approximating matrix multiplication. Further\
, this decomposition may have additional gains in a quantum computation. T
he talk is designed to introduce this matrix decomposition to students.\n\
n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:68@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240812T143000
DTEND;TZID=Asia/Kolkata:20240812T153000
DTSTAMP:20240730T054550Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-0230-august-a-new-locall
y-adaptive-nonparametric-regression-method/
SUMMARY:{Seminar} @ CDS: #102: 02:30 August: "A New Locally Adaptive Nonpar
ametric Regression Method."
DESCRIPTION:Department of Computational and Data Sciences\nDepartment Semin
ar\n\n\n\nSpeaker : Prof. Sabyasachi Chatterjee\, University of Illinois a
t Urbana Champaign.\nTitle : A New Locally Adaptive Nonparametric Regressi
on Method.\nDate &\; Time : August 12\, 2024\, 2:30 PM\nVenue : # 102\,
CDS Seminar Hall\n\n\n\nABSTRACT\n\nWe propose and study a new locally ad
aptive nonparametric regression method. The method performs variable bandw
idth local averaging/local polynomial regression. To certify its local ada
ptivity we show that it adapts near optimally to the local Holder smoothne
ss exponent of the regression function at any point in the domain. Despite
the vast literature on Nonparametric Regression\, we only know one existi
ng method which attains such local adaptivity proveably. This method is kn
own as Lepski's method. There are some drawbacks to Lepski's method such a
s a) it is specifically tailored to a given class of functions such as Hol
der Smooth function class b) it is rather theoretical and impractical to i
mplement with effectively many tuning parameters. Our proposed method seem
s to overcome these drawbacks. Firstly\, our method is defined without any
reference to any function class and secondly there is only one tuning par
ameter\, which when set properly\, adjusts all the bandwidths at all locat
ions near optimally. Our method is practically implementable and appears t
o perform reasonably well in our numerical experiments.\n\nBIO: Dr. Sabyas
achi Chatterjee is currently working as an Associate Professor in the Depa
rtment of Statistics at the University of Illinois at Urbana Champaign. He
earned his PhD in Statistics at Yale University\, USA. Then\, he worked a
s a Postdoctoral Research Scholar and William Kruskal Instructor in the De
partment of Statistics at the University of Chicago\, USA. His research in
terests lie in Mathematical Statistics and Machine Learning Theory\, along
with their connections with Information Theory\, Probability\, and Optimi
zation. For more about his research\, please see https://sabyasachi.web.il
linois.edu/\n\nHost Faculty: Dr. Ratikanta Behera\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:70@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240812T110000
DTEND;TZID=Asia/Kolkata:20240812T120000
DTSTAMP:20240805T114329Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-cds-12-august-2024-e
fficient-and-effective-algorithms-for-improving-the-robustness-of-deep-neu
ral-networks/
SUMMARY:Ph.D. Thesis Defense: CDS: 12\, August 2024 "Efficient and Effectiv
e Algorithms for Improving the Robustness of Deep Neural Networks"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Def
ense\n\n\n\nSpeaker : Ms. Sravanti Addepalli\nS.R. Number : 06-18-02-17-12
-18-1-15587\nTitle : "Efficient and Effective Algorithms for Improving the
Robustness of Deep Neural Networks"\nThesis examiner: Prof. Vineeth Balas
ubramanian\, IIT Hyderabad\nResearch Supervisor: Prof. Venkatesh Babu R
\nDate &\; Time : August 12\, 2024 (Monday) at 11:00 AM\nVenue : # 102
CDS Seminar Hall\n\n\n\nABSTRACT\nDeep neural networks (DNNs) have achieve
d remarkable success across various domains\, yet their vulnerability to a
dversarial attacks and distribution shifts remains a significant challenge
. This thesis presents novel methodologies to enhance DNN robustness\, foc
using on efficiency\, effectiveness\, and practical applicability.\n\nThe
first part of the thesis concentrates on developing computationally effici
ent adversarial defenses. Traditional adversarial training methods are oft
en computationally intensive due to the generation of adversarial examples
through multiple optimization steps. To address this\, we introduce Bit P
lane Feature Consistency (BPFC)\, a regularizer that promotes robustness w
ithout requiring adversarial examples during training. Furthermore\, we pr
opose Guided Adversarial Training (GAT) and Nuclear Norm Adversarial Train
ing (NuAT) to mitigate the gradient masking issue prevalent in single-step
adversarial training\, leading to improved robustness without sacrificing
computational efficiency.\n\nThe second part focuses on improving the eff
ectiveness of adversarial training. While adversarial training enhances ro
bustness\, it comes at the cost of reduced accuracy on clean data. To addr
ess this\, we introduce Feature Level Stochastic Smoothing (FLSS)\, a meth
od that combines adversarial training with detection to boost robustness a
nd accuracy. Additionally\, we propose Oracle-Aligned Adversarial Training
(OAAT) to address the robustness-accuracy trade-off at large perturbation
bounds. To further enhance adversarial training\, we explore the integrat
ion of data augmentation techniques through Diverse Augmentation based Joi
nt Adversarial Training (DAJAT).\n\nThe third part of the thesis focuses o
n improving the efficiency and effectiveness of self-supervised training f
or robust representation learning. We investigate the potential of combini
ng the popular instance-discrimination task with auxiliary tasks such as r
otation prediction to reduce noise in the training objective and improve t
he quality of learned representations. We further utilize these self-super
vised pretrained models in a teacher-student distillation setting for trai
ning adversarially robust models without labels using the proposed method
Projected Feature Adversarial Training (ProFeAT).\n\nThe final part of the
thesis addresses the brittleness of DNNs to distribution shifts. We propo
se the Feature Replication Hypothesis (FRH) to explain the underlying caus
es of vulnerability to distribution shifts. To mitigate this\, we introduc
e the Feature Reconstruction Regularizer (FRR) that encourages the learnin
g of diverse feature representations. Additionally\, Diversify-Aggregate-R
epeat Training (DART) is proposed to improve generalization of DNNs by tra
ining diverse models in parallel\, and aggregating their weights intermitt
ently over training. We finally propose Vision-Language to Vision - Align\
, Distill\, Predict (VL2V-ADiP)\, a teacher-student setting to utilize the
superior generalization of Vision-Language Models (VLMs) for improving th
e OOD generalization in vision tasks.\n\nThrough these contributions\, thi
s thesis advances the state-of-the-art in DNN robustness by providing prac
tical and effective solutions to address the challenges posed by adversari
al attacks and distribution shifts. The proposed methods demonstrate signi
ficant improvements in both robustness and accuracy\, paving the way for m
ore reliable and resilient models.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:71@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240809T143000
DTEND;TZID=Asia/Kolkata:20240809T153000
DTSTAMP:20240808T055937Z
URL:https://cds.iisc.ac.in/events/interactive-talk-on-how-fast-can-we-comp
ute-with-special-mentorship-session/
SUMMARY:Interactive Talk on "How Fast Can We Compute?" with Special Mentors
hip Session
DESCRIPTION:Greetings from the IISc ACM Student Chapter!\n\nWe are thrilled
to share that Prof. Sajal K. Das\, a renowned computer scientist\, IEEE F
ellow\, and esteemed alumnus of IISc\, is visiting our campus as the Infos
ys Visiting Chair Professor. During his visit\, you will have the opportun
ity to engage with him through an intriguing talk. Renowned for his except
ional mentorship and guidance\, Prof. Das will also host a special interac
tive mentorship session\, which will be followed by high tea. This is a wo
nderful chance to gain insights from a distinguished expert and connect wi
th him personally.\n\nWe request you to express your interest by filling o
ut this form (https://forms.office.com/r/zUf346NxW2). We look forward to y
our enthusiastic response. Please find the abstract and details of the tal
k below.\n\n\n\nSpeaker: Prof. Sajal K. Das\, Missouri University of Scien
ce and Technology\nTitle: How Fast Can We Compute?\nDate/Time: 9th August
\, 2024 (Friday) 2:30 PM - 3:30 PM\nVenue: CDS 102 Seminar Hall\n\n\n\nAb
stract: What is Computational Thinking? Do you apply it efficiently and ef
fectively in daily life? Do you watch movies\, read mystery books\, play a
ny sports or games? Do you find beauty and joy in learning and problem sol
ving? Are you motivated? What makes you happy? What is common among passio
nate researchers\, story tellers\, movie directors\, and sports enthusiast
s? This talk is an attempt to unfold the mystery by picking simple computa
tional or recreational problems and solve them fast with elegance and beau
ty.\n\nBio: Dr. Sajal K. Das is a Curators’ Distinguished Professor of C
omputer Science and Daniel St. Clair Endowed Chair at Missouri University
of Science and Technology\, Rolla\, USA where he was the Chair of the Com
puter Science Department during 2013-2017. Previously\, he served the US N
ational Science Foundation as a Program Director in computer science. His
interdisciplinary research interests include cyber-physical systems (CPS)\
, IoT\, drones\, cybersecurity\, machine learning\, data science\, wireles
s and sensor networks\, mobile and pervasive computing\, smart environment
s\, edge/cloud computing\, and applied graph theory. He has made fundamen
tal contributions to these areas and published extensively (more than 600
papers) in high quality journals and peer-reviewed conference proceedings\
, 59 book chapters\, 4 books\, and 5 US patents. Dr. Das has directed over
$24 million funded research projects. According to Google Scholar\, his h
-index is 100 with more than 42\,500 citations. He is the founding Editor-
in-Chief of Elsevier’s Pervasive and Mobile Computing journal and serves
as an Associate Editor of the IEEE Transactions on Dependable and Secure
Computing\, IEEE Transactions on Mobile Computing\, IEEE Transactions on S
ustainable Computing\, ACM/IEEE Transactions on Networking\, and ACM Trans
actions on Sensor Networks. A founder of conferences like IEEE PerCom\, IE
EE WoWMoM\, IEEE SMARTCOMP\, and ACM ICDCN\, he has served as the General
and Technical Program Chair of numerous conferences. He is a recipient of
12 Best Paper Awards at prestigious conferences including ACM MobiCom\, IE
EE PerCom\, and IEEE SMARTCOMP. He also received awards of excellence for
teaching\, mentoring\, research\, and innovation including the IEEE Comput
er Society’s Technical Achievement award and the University of Missouri
System President’s Award for Sustained Career Excellence. He has graduat
ed 11 postdoctoral fellows\, 51 Ph.D. students\, and 31 MS thesis and nume
rous undergraduate research students. A Distinguished Alumnus of the IISc
and an Infosys Visiting Chair Professor at IISc\, Dr. Das a Fellow of the
IEEE\, National Academy of Inventors (NAI)\, and Asia-Pacific Artificial I
ntelligence Association (AAIA).\n\nRegards\,\nIISc ACM Student Chapter
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:64@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240805T150000
DTEND;TZID=Asia/Kolkata:20240805T160000
DTSTAMP:20240716T030555Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-cds-05-august-2024-a
-scalable-asynchronous-discontinuous-galerkin-method-for-massively-paralle
l-flow-simulations/
SUMMARY:Ph.D. Thesis Defense: CDS: 05\, August 2024 "A scalable asynchronou
s discontinuous Galerkin method for massively parallel flow simulations"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Def
ense\n\n\n\nSpeaker : Mr. Shubham Kumar Goswami\nS.R. Number : 06-18-00-10
-12-19-1-17224\nTitle : "A scalable asynchronous discontinuous Galerkin me
thod for massively parallel flow simulations "\nThesis examiner: Dr. Prave
en Chandrashekar\, Center for Applicable Mathematics Tata Institute of Fun
damental Research.\nResearch Supervisor: Dr. Konduri Aditya\nDate &\; T
ime : August 05\, 2024 (Monday) at 03:00 PM\nVenue : # 102 CDS Seminar Hal
l\n\n\n\nABSTRACT\n\nAccurate simulations of turbulent flows are crucial f
or understanding numerous complex phenomena in engineered systems and natu
ral processes. Notably\, under realistic conditions with high Reynolds num
bers and complex geometries\, the partial differential equations (PDEs) go
verning these fluid flows are highly nonlinear and are solved numerically
using PDE solvers. Due to the presence of multiple length and time scales
inherent to turbulent flows\, these simulations are often computationally
expensive\, necessitating the use of massively parallel supercomputers. De
spite several advancements in the development of scalable PDE solvers\, th
ey face scalability challenges at extreme scales due to communication over
head. To address this issue\, an asynchronous computing approach that rela
xes communication/synchronization at a mathematical level has been develop
ed with finite difference schemes. However\, these schemes are not amenabl
e to capture flows in complex geometries with unstructured meshes. The obj
ective of this thesis is to develop an asynchronous discontinuous Galerkin
(ADG) method with the potential to provide high-order accurate solutions
for various flow problems on structured and unstructured meshes and demons
trate its scalability. The thesis includes developing an approach to coupl
e asynchronous schemes with low-storage Runge-Kutta schemes\, then introdu
cing the ADG method and investigating its properties\, and finally impleme
nting the proposed method into deal.II (open-source library) for scalabili
ty demonstrations.\n\nBased on the asynchronous computing approach\, sever
al PDE solvers have been developed that use high-order asynchrony-tolerant
(AT) finite difference schemes for spatial discretization to simulate rea
cting and non-reacting turbulent flows\, achieving significant improvement
s in scalability. For time integration\, they use either multi-step Adams-
Bashforth schemes\, which possess poor stability\, or multi-stage Runge-Ku
tta (RK) schemes with an over-decomposed domain that necessitates larger m
essage sizes for communication and redundant computations. In this work\,
we propose a novel method to couple asynchrony-tolerant and low-storage ex
plicit RK (LSERK) schemes to solve time-dependent PDEs with reduced commun
ication efforts. We develop new schemes for ghost or buffer point updates
that are necessary to maintain the desired order of accuracy. The accuracy
of this method has been investigated both theoretically and numerically u
sing simple one-dimensional linear model equations. Thereafter\, we demons
trate its scalability through three-dimensional simulations of decaying Bu
rgers’ turbulence performed using two different asynchronous algorithms:
communication-avoiding and synchronization-avoiding algorithms. Scalabili
ty studies up to 27\,000 cores yielded a speed-up of up to 6x compared to
a baseline synchronous algorithm.\n\nIn recent years\, the discontinuous G
alerkin (DG) method has gained considerable attention in developing PDE so
lvers\, particularly for nonlinear hyperbolic problems\, due to its abilit
y to provide high-order accurate solutions in complex geometries\, capture
discontinuities\, and exhibit high arithmetic intensity. However\, the sc
alability of DG-based solvers is hindered by communication bottlenecks tha
t arise at extreme scales. In this work\, we introduce the asynchronous DG
(ADG) method\, which combines the benefits of the DG method with asynchro
nous computing by relaxing the need for data communication and synchroniza
tion at the mathematical level. The proposed ADG method ensures local cons
ervation and effectively addresses challenges arising from asynchrony. To
assess its stability\, we employ Fourier-mode analysis to examine the diss
ipation and dispersion behavior of fully-discrete DG and ADG schemes with
the Runge-Kutta (RK) time integration schemes across a wide range of waven
umbers. Furthermore\, we present an error analysis demonstrating that the
ADG method with standard numerical fluxes achieves at most first-order acc
uracy. To recover accuracy\, we derived asynchrony-tolerant (AT) fluxes th
at utilize data from multiple time levels. Finally\, extensive numerical e
xperiments are conducted to validate the performance and accuracy of the A
DG-AT scheme for both linear and nonlinear problems.\n\nWith the developme
nt of the asynchronous discontinuous Galerkin (ADG) method\, we finally pu
t our focus on implementing and evaluating its performance in solving hype
rbolic equations with shocks/discontinuities.\n\nTo achieve this\, we chos
e a highly scalable DG solver for compressible Euler equations from deal.I
I\, which is one of the widely used open-source finite element libraries.
The solver uses low-storage explicit Runge-Kutta schemes for the time inte
gration. We implemented the ADG method in deal.II\, incorporating the comm
unication-avoiding algorithm (CAA)\, and performed accuracy validation and
scalability benchmarks. The results showcase the accuracy limitations of
standard ADG schemes and the effectiveness of newly developed asynchrony-t
olerant (AT) fluxes. Strong scaling results are provided for both synchron
ous and asynchronous DG solvers\, demonstrating a speedup of up to 80% wit
h the ADG method at an extreme scale with 9216 cores.\n\nThis thesis focus
ed on the development of scalable PDE solvers based on the asynchronous di
scontinuous Galerkin method for massively parallel flow simulations. Altho
ugh these advancements were specifically geared towards the DG method\, th
ey are also applicable to the finite volume (FV) method and can be easily
integrated into commercial FV-based PDE solvers. The overall work highligh
ts the potential benefits of the asynchronous approach for the development
of accurate and scalable DG and FV-based PDE solvers\, paving the way for
simulations of complex physical systems on massively parallel supercomput
ers.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:69@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240805T033000
DTEND;TZID=Asia/Kolkata:20240805T043000
DTSTAMP:20240802T084826Z
URL:https://cds.iisc.ac.in/events/seminar-serc-4th-floor-auditorium-05th-a
ugust-protein-sequence-annotation-using-language-models/
SUMMARY:{Seminar} @ SERC : 4th Floor Auditorium : 05th August: "Protein Seq
uence Annotation using Language Models"
DESCRIPTION:We welcome you to CDS-KIAC talk on 05 August 2024 (Monday). The
details are as below:\n\n\n\nSpeaker: Dr. Kumaresh\, post doctoral resear
cher\, Eddy Lab at Harvard University.\nTitle: "Protein Sequence Annotatio
n using Language Models"\nDate and Time: August 05\, 2024\, 3:30 PM\nVenue
: #421\, SERC Auditorium\n\n\n\nAbstract:\nProtein function inference reli
es on annotating protein domains via sequence similarity\, often modeled t
hrough profile Hidden Markov Models (profile HMMs)\, which capture evoluti
onary diversity within related domains. However\, profile HMMs make strong
simplifying independence assumptions when modeling residues in a sequence
. Here\, we introduce PSALM (Protein Sequence Annotation with Language Mod
els)\, a hierarchical approach that relaxes these assumptions and uses rep
resentations of protein sequences learned by protein language models to en
able high-sensitivity\, high-specificity residue-level protein sequence an
notation. We validate PSALM’s performance on a curated set of "ground tr
uth" annotations determined by a profile HMM-based method and highlight PS
ALM as a promising alternative for protein sequence annotation.\n\nBio of
Speaker:\nKumaresh is a post doctoral researcher in the Eddy Lab at Harvar
d University working on machine learning models for annotating\, understan
ding and analyzing protein sequences. He has a PhD from Harvard University
where he worked in systems neuroscience\, building models of decision mak
ing and attentional switching using zebrafish as a model organism. Kumares
h's undergraduate and Masters training is in Computer Science and Electric
al Engineering from IIIT Bangalore and he brings this strong computational
background to tackle complex real world biological problems.\n\n\n\nALL A
RE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:67@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240802T113000
DTEND;TZID=Asia/Kolkata:20240802T123000
DTSTAMP:20240725T145159Z
URL:https://cds.iisc.ac.in/events/mtech-research-thesis-defense-cds-02-aug
ust-2024-scalable-video-data-management-and-visual-querying-for-autonomous
-camera-networks/
SUMMARY:Mtech Research Thesis Defense: CDS: 02\, August 2024 "Scalable Vide
o Data Management and Visual Querying for Autonomous Camera Networks"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nMtech Research T
hesis Defense\n\n\n\nSpeaker : Ms. Bharati Khanijo\nS.R. Number : 06-18-02
-10-12-19-1-17219\nTitle : "Scalable Video Data Management and Visual Quer
ying for Autonomous Camera Networks"\nThesis examiner : Dr. Soma Biswas\nR
esearch Supervisor: Dr. Yogesh Simmhan\nDate &\; Time : August 02\, 202
4 (Friday) at 11:30 AM\nVenue : # 202 CDS Classroom\n\n\n\nABSTRACT\n\nVid
eo data has been historically known not only for its unstructured nature a
nd rich semantic content but also for scalability issues in terms of stora
ge and analytics. Mobile aerial platforms like drones capture such videos
across space and time. Advances in computer vision and deep learning enabl
e automatic extraction of rich semantic information from video data\, lead
ing to applications where the stored video data can be used to study and a
nalyze the world retrospectively and automatically. However\, recent resea
rch has highlighted the compute-intensive nature of such Deep Neural Netwo
rk (DNN) models\, e.g.\, for accurate object detection\, leading to high c
omputing costs that limit their applicability for brute-force analysis of
all historical videos. Also\, an efficient design of such applications oft
en requires co-analysis of video data along with associated geospatial and
temporal metadata\, which is a challenge.\n\nWe propose a geospatial-temp
oral video query system with support for semantic queries for drone videos
\, extending an existing spatial-temporal database and contemporary object
detection models. We develop a heuristic to enable better reuse of semant
ic object detection results obtained from different model configurations (
eg. object detection model and its input resolution) . The system further
motivates the need for optimizations for retrospective semantic analysis a
nd storage for drone videos\, which is addressed by our novel DDownscale m
ethod and the associated ingest pipeline.\n\nPrior optimizations on semant
ic querying over video data focus on static cameras from city-scale traffi
c/surveillance camera networks\, often leveraging the spatial and temporal
characteristics of associated videos\, which are absent in videos recorde
d by mobile drone cameras. We specifically focus on two such characteristi
cs of drone videos. One is that drone videos have shorter durations\, unli
ke those captured by static cameras. Another is that there can be large va
riations in the level of detail of information captured across a fleet of
drone cameras due to differences in the resolution of the camera\, the alt
itude\, and the orientation from which the videos were captured.\n\nSpecif
ically\, we address the need to intelligently scale-down the spatial resol
ution of videos to reduce the video storage costs and semantic query/infer
encing time. However\, conventional methods of manual or profiling-based e
stimation of the ideal scaling ratio are compute-intensive and/or time-con
suming for such heterogeneous feeds. We propose DDownscale\, a novel metho
d to dynamically select the downscale factor for a video by utilizing the
information on the object size in the video. We model the downscale factor
and associated drop in relative recall due to downscaling as a function o
f object size in the downscaled video and demonstrated that for a given DN
N model and class of interest\, DDownscale generalizes well to the evaluat
ed datasets. A DDownscale inequality between the relative recall drop and
the hyperparameters of the method is derived. This satisfies ~ 98% of the
dynamically downscaled videos across real world video datasets\, objects o
f interest and parameters. The algorithm achieves over 25% reduction in to
tal object detection time and 31% reduction in storage on average compared
to the baseline of storing/inferencing at the original resolution\, for d
ifferent user-specified target reduction in recall values ranging from 1 -
- 30%\, and 96% of the downscaled videos are within the target recall drop
\, for the evaluated datasets and object detection models.\n\nA simpler sp
ecification at the time of ingest of target level of detail (average groun
d spatial distance) captured in the video and the harmonic mean of relativ
e recall drop for the class of smallest object of interest and selected ob
ject detection model was derived using the above modeling to aid in the se
lection of a target level of detail. Additionally\, we develop an ingest p
ipeline that reduces the time to ingest drone videos using this dynamicall
y downscaling method over heterogeneous edge accelerators\, and reduce the
average turnaround time to ingest data from multiple clients by ~ 66%\, d
espite the downscaling time overhead\, compared to uploading original reso
lution video without downscaling.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:61@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata;VALUE=DATE:20240801
DTEND;TZID=Asia/Kolkata;VALUE=DATE:20240803
DTSTAMP:20240704T124405Z
URL:https://cds.iisc.ac.in/events/workshop-on-artificial-intelligence-in-p
recision-medicine-2/
SUMMARY:Workshop on Artificial Intelligence in Precision Medicine
DESCRIPTION:Dear All\,\n\nWe are pleased to announce a two-day long:\n\nWOR
KSHOP ON ARTIFICIAL INTELLIGENCE IN PRECISION MEDICINE\n\nDate: 1-2 Augu
st\, 2024\n\nLocation: #102-Seminar Hall\, Computational &\; Data Scie
nces Department\, IISc\n\nThe workshop aims to foster a deeper understan
ding of ongoing research in both Industry and Academia by bringing togethe
r leaders from both. We hope to provide a comprehensive overview of recent
advancements in precision medicine\, and also explore its wide-ranging ap
plications.\n\nFor more information\, please visit our website: website\n\
nMandatory Poster Submission\n\nIn line with our commitment to fostering i
nteractive discussions and knowledge exchange\, we require all attendees t
o submit posters presenting their research contributions relevant to the
workshop themes. Poster submissions are mandatory for participation in th
e workshop.\n\nYou can submit your poster and register yourself here: reg
istration form\n\nNote: All the participants are requested to kindly carry
your own laptops for the hands-on session.\n\nWarm Regards\,\nOrganising
Team\n\n\n\n
CATEGORIES:Events
END:VEVENT
BEGIN:VEVENT
UID:65@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240725T100000
DTEND;TZID=Asia/Kolkata:20240725T110000
DTSTAMP:20240716T030922Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-cds-25-july-2024-fas
t-and-scalable-algorithms-for-intelligent-routing-of-autonomous-marine-veh
icles/
SUMMARY:Ph.D. Thesis Defense: CDS: 25\, July 2024 "Fast and Scalable Algori
thms for Intelligent Routing of Autonomous Marine Vehicles"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Def
ense\n\n\n\nSpeaker : Mr. Rohit Chowdhury\nS.R. Number : 06-18-01-10-12-18
-1-16320\nTitle : "Fast and Scalable Algorithms for Intelligent Routing of
Autonomous Marine Vehicles"\nThesis examiner: Dr. Indranil Saha\, IIT Kan
pur\nResearch Supervisor: Dr. Deepak Subramani\nDate &\; Time : July 25
\, 2024 (Thursday) at 10:00 AM\nVenue : # 102 CDS Seminar Hall\n\n\n\nABST
RACT\n\nAutonomous marine agents play a pivotal role in diverse ocean appl
ications. These agents serve as indispensable instruments for acquiring cr
ucial environmental information. They are used to explore and monitor of h
arsh environments\, e.g.\, to map ocean topography\, study coral reefs\, s
earch and rescue\, structural monitoring of oil and gas installations etc.
In naval security\, these agents are used for surveillance and strategic
monitoring of maritime activities. Building intelligence to optimally use
these agents is essential for reducing operational costs.\n\nThe path plan
ning problem is as follows. An autonomous marine agent must optimally trav
erse from a given start location to a given target location in a stochasti
c dynamic velocity field like ocean currents while avoiding obstacles or r
estricted regions in the flow. A key challenge is that the agent is heavil
y advected by the flow. The optimality may refer to minimising expected tr
avel time or energy consumption\, data collection or risk of failure. Whil
e there are multiple methods of solving path planning problems\, each with
its challenges\, we develop and use a fast and scalable MDP-based offline
planning software that computes optimal policies\, and a novel sequence-m
odelling-based deep learning approach for onboard routing and dynamic plan
ning\, where the objective is to learn optimal action sequences for the ag
ent. The goal of this thesis is to develop efficient\, fast and scalable A
rtificial intelligence algorithms for optimal planning and on-board routin
g algorithms for autonomous marine agents in stochastic dynamic environmen
ts.\n\nThe thesis comprises five works organised into two parts based on t
he solution approach. In the first part\, we model the path planning probl
em as a Markov Decision Process (MDP) and aim to compute an optimal policy
. However\, the key challenge here is that solving an MDP can be prohibiti
vely expensive for large state and action spaces. To overcome this challen
ge\, we either approximate the optimal policy or accelerate the computatio
n using GPUs.\n\n Physics-driven Q-learning for onboard routing: First\,
the distribution of exact time-optimal paths predicted by stochastic Dynam
ically Orthogonal (DO) Hamilton-Jacobi level set partial differential equa
tions (HJLS PDEs) are utilised to learn an initial action-value function t
hat approximates the optimal policy. The flow data collected by onboard se
nsors are utilised to get a posterior estimate of the environment. The app
roximated optimal policy is refined in-mission by performing epsilon greed
y Q-learning in simulated posterior environments. We showcase the computat
ional advantage of the approach at the cost of approximating the optimal p
olicy.\n GPU-accelerated path planning: We compute an exact optimal polic
y by solving the path planning problem modelled as an MDP. To solve large-
scale real-time problems\, which can otherwise be computationally expensiv
e\, we introduce an efficient end-to-end GPU accelerated algorithm that bu
ilds the MDP model (computing transition probabilities and expected one-st
ep rewards) and solves the MDP to compute an optimal policy. We develop me
thodical and algorithmic solutions to overcome the limited global memory o
f GPUs by using a dynamic reduced-order representation of the ocean flows\
, leveraging the sparse nature of the state transition probability matrix
and introducing a neighbouring subgrid concept to save memory and reduce t
he computational effort. We achieve significant speedups compared to conve
ntional sequential computation.\n Multi-objective GPU-accelerated path pl
anning: The end-to-end GPU accelerated MDP solver is extended to a multi-o
bjective path planner to solve multi-objective optimisation problems in pa
th planning\, like minimising both the expected mission completion time an
d energy consumption. MDPs are modelled with scalarised rewards for multip
le objectives. The solver is used to solve numerous instances of complex s
cenarios with other sources of uncertainty in the environment\, enabling u
s to compute optimal operating curves in a fraction of the time compared t
o traditional solvers.\n\nIn the second part\, we convert the optimal path
planning problem into a supervised learning problem via sequence modellin
g. This approach involves learning optimal action sequences based on the a
vailable environment information and expert trajectories. It also avoids t
he issue of re-computing optimal policies for onboard routing.\n Intellig
ent onboard routing using decision transformers: We develop a novel\, deep
learning method based on the decision transformer (decoder-only model) fo
r onboard routing of autonomous marine agents. Training data is obtained f
rom aforementioned HJLS PDE or MDP solvers\, which is further processed to
sequences of states\, actions and returns. The model is autoregressively
trained on these sequences and then tested in different environment settin
gs. We demonstrate that (i) a trained agent learns to infer the surroundin
g flow and perform optimal onboard routing when the agent's state estimati
on is accurate\,(ii) specifying the target locations (in case of multiple
targets) as a part of the state enables a trained agent to route itself to
the correct destination\, and (iii) a trained agent is robust to limited
noise in state transitions and is capable of reaching target locations in
completely new flow scenarios. We extensively showcase end-to-end planning
and onboard routing in various canonical and idealised ocean flow scenari
os.\n Path planning with environment encoders and action decoders: We pro
pose a novel combination of dynamically orthogonal flow representation wit
h uncertainty and a transformer model (encoder-decoder) for the path plann
ing task. We model the problem as a sequence-to-sequence translation task
where the source sequence is the agent's knowledge representation of the u
ncertain environmental flow. The target sequence is the optimal sequence o
f actions the agent must execute. We demonstrate that a trained transforme
r model can predict near-optimal paths for unseen flow realisations and ob
stacle configurations in a fraction of the time required by traditional pl
anners. Validation is performed to show generalisation in unseen obstacle
configurations. We also analyse the predictions of both transformer models
\, viz\, decoder only and encoder-decoder and explain the inner mechanics
of learning through a novel visualisation of self-attention of actions and
states on the trajectories.\n\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:66@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240718T113000
DTEND;TZID=Asia/Kolkata:20240718T123000
DTSTAMP:20240716T045953Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-18th-july-scene-understa
nding-for-safe-and-autonomous-navigation/
SUMMARY:{Seminar} @ CDS: #102 : 18th July: "Scene Understanding for Safe an
d Autonomous Navigation"
DESCRIPTION:We welcome you to CDS-KIAC talk on 18 July 2024 (Thursday). The
details are as below:\n\n\n\nSpeaker: Prof. Amit K. Roy-Chowdhury\, Unive
rsity of California\nTitle:Scene Understanding for Safe and Autonomous Nav
igation\nDate and Time: July 18\, 2024\, 11:30 AM\nVenue: #102\, CDS Semin
ar Hall.\n\n\n\nAbstract:\nAutonomous navigation remains one of the most c
hallenging problems in intelligent systems largely because of the close in
tegration of scene understanding and planning that needs to happen. The sc
ene understanding requires analysis of objects and their collections acros
s various scale\, from individual people and their actions to wide-area an
alysis that could span the interactions of these people with many other ob
jects in the scene. An integrated view that is able to span across these r
anges of scale is necessary for robust decision making. In this talk\, we
will consider a variety of scene understanding problems that need to be so
lved for autonomous navigation to be successful. At the level of individua
l people\, we will show how to estimate the pose of each individual person
under challenging real-life conditions like significant occlusions. At th
e next higher scale when there are interactions among small groups of indi
viduals and objects\, we will demonstrate the power of scene graphs to mod
el the semantics of the scene. At a yet higher level\, we will show how to
track objects across non-overlapping cameras spread over large areas. Rob
ustness to a variety of operational domains will be considered through all
of these tasks. In spite of this\, it is unlikely that perfect scene unde
rstanding will be achieved and any autonomous agent will need to occasiona
lly interact with human experts\; we show how this can be achieved with na
tural language feedback leveraging upon the power of recently developed vi
sion-language models.\n\nBio of Speaker:\nAmit Roy-Chowdhury received his
PhD from the University of Maryland\, College Park (UMCP) in 2002 and join
ed the University of California\, Riverside (UCR) in 2004 where he is a Pr
ofessor and Bourns Family Faculty Fellow of Electrical and Computer Engine
ering\, Cooperating Faculty in Computer Science and Engineering\, and Dire
ctor of the Center for Robotics and Intelligent Systems. He leads the Vide
o Computing Group at UCR\, working on foundational principles of computer
vision\, image processing\, and machine learning\, with applications in cy
ber-physical\, autonomous and intelligent systems. He has published over 2
00 papers in peer-reviewed journals and conferences. He has published two
monographs: Camera Networks: The Acquisition and Analysis of Videos Over W
ide Areas and Person Re-identification with Limited Supervision. He is on
the editorial boards of major journals and program committees of the main
conferences in his area. He is a Fellow of the IEEE and IAPR\, received th
e Doctoral Dissertation Advising/Mentoring Award from UCR\, and the ECE Di
stinguished Alumni Award from UMCP.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:63@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240718T093000
DTEND;TZID=Asia/Kolkata:20240718T103000
DTSTAMP:20240710T152647Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-colloquium-cdslea
rning-multiple-initial-conditions-using-physics-informed-neural-networks-p
inns/
SUMMARY:M.Tech Research Thesis {Colloquium}: CDS:"Learning Multiple Initial
Conditions Using Physics-Informed Neural Networks (PINNs)"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Resarch T
hesis Colloquium\n\n\n\nSpeaker : Mr. Mahesh Tom\nS.R. Number : 06-18-01-1
0-22-21-1-20317\nTitle : "Learning Multiple Initial Conditions Using Physi
cs-Informed Neural Networks (PINNs)"\nResearch Supervisor : Prof. Sashikum
aar Ganesan\nDate &\; Time : July 18\, 2024 (Thursday)\, 9:30 AM\nVenue
: # 102 CDS Seminar Hall\n\nABSTRACT\n\nPhysics-Informed Neural Networks
(PINNs) and their variants have emerged as tools for solving differential
equations in the past few years. Although several variants of PINNs have b
een proposed for time-dependent partial differential equations (PDEs)\, th
e majority of these physics-informed approaches are based on solving a pro
blem for a single set of initial conditions. In this work\, we consider on
e-dimensional time-dependent PDEs and focus on solving multiple initial co
nditions (ICs) with a single network simultaneously. Trying to solve multi
ple ICs in a single network presents certain challenges\, such as spectral
bias\, that we address in our work. We also look at how our approach perf
orms in the FastVPINNs framework to solve multiple ICs using Variational P
hysics-Informed Neural Networks (VPINNs). The choice of activation functio
ns is crucial in the performance of a network\; hence we also test the inf
luence of various activation functions on FastVPINNs for some standard tes
t cases. While training multiple ICs\, we also look at the impact of the n
etwork parameters and how they contribute to each trained task via an abla
tion study.\n\nOnce we have a fully trained model that works on multiple I
Cs\, incorporating new ICs without having to retrain all the previous ICs
is a challenging task\, and a brute-force way of training all the ICs agai
n is not always feasible. To this end\, we explore the usage of elastic we
ight consolidation (EWC)\, a regularization technique that is used in cont
inual learning\, and study its effect on PINNs for training new ICs.\n\n\n
\nALL ARE WELCOME
CATEGORIES:Events,MTech Research Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:62@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240716T110000
DTEND;TZID=Asia/Kolkata:20240716T120000
DTSTAMP:20240708T063449Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-102-cds-16-july-2
024-structural-and-functional-studies-on-the-hypothetical-protein-ttha1873
-from-thermus-thermophilus/
SUMMARY:Ph.D: Thesis Colloquium: 102 : CDS: 16\, July 2024 "Structural and
functional studies on the hypothetical protein TTHA1873 from Thermus therm
ophilus"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Col
loquium\n\n\n\nSpeaker : Mr. Yuvaraj I\nS.R. Number : 06-18-01-10-12-13-1-
10377\nTitle : "Structural and functional studies on the hypothetical prot
ein TTHA1873 from Thermus thermophilus"\nResearch Supervisor : Prof. K. Se
kar\nDate &\; Time : July 16\, 2024 (Tuesday)\, 11:00 AM\nVenue : # 102
CDS Seminar Hall\n\n\n\nABSTRACT\n\nThis thesis reports a detailed study
on the structural and functional characterization of the hypothetical prot
ein (TTHA1873) from Thermus thermophilus. In this study\, we characterized
both the structure and function of TTHA1873. To elucidate the novel struc
ture of this uncharacterized protein\, the study employed a heavy atom der
ivative compound K 2 [HgI 4 ] to obtain an anomalous signal using home sou
rce X-ray. The structural information obtained was then used to infer the
function of this protein. Biochemical experiments demonstrated that TTHA18
73 acts as a nuclease\, indiscriminately cutting methylated and non-methyl
ated DNA in divalent metal ions and relaxing plasmid DNA in the presence o
f ATP. Its activity is inhibited by EDTA. Structural analysis identified f
unctionally important residues\, and molecular dynamics simulations were c
onducted to investigate the effects of mutating two critical residues on D
NA binding.\n\nThe study explored the effects of high temperatures on this
protein through molecular dynamics\, identifying temperature-sensitive re
gions and providing insights into thermal denaturation. The use of heavy a
tom compound K 2 [HgI 4 ] to obtain phases has been previously reported in
the literature. Still\, the protocols used to solve the three-dimensional
structure of this protein may be particularly useful in challenging cases
where molecular replacement is ineffective or when no similar structures
are available in the Protein Data Bank (PDB)\, without relying on a synchr
otron source.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:58@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240708T113000
DTEND;TZID=Asia/Kolkata:20240708T123000
DTSTAMP:20240619T064050Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-defense-cds-08-ju
ly-2024-sequence-alignment-to-cyclic-pangenome-graphs/
SUMMARY:M.Tech Research: Thesis Defense: CDS: 08\, July 2024 "Sequence Alig
nment to Cyclic Pangenome Graphs"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\n\nM.Tech Resea
rch Thesis Defense\n\n\n\n\nSpeaker : Ms. Jyotsh
na Rajput\nS.R. Number : 06-18-01-10-22-21-1-19943\nTitle
: "Sequence Alignment to Cyclic Pang
enome Graphs"\nThesis examiner : Prof. L Sunil Chandran\, Dept. of Comp
uter Science and Automation\, IISc.\nResearch Supervisor: Dr. Chirag Jai
n\nDate &\; Time : July 08\, 2024 (Monday) at 11:30 AM\nV
enue : CDS # 308\n\n\n\n\n\nABSTRACT\n\
n\nThe growing availability of genome sequences of several species\, inclu
ding humans\, has created the opportunity to utilize multiple reference ge
nomes for bioinformatics analyses and improve the accuracy of genome reseq
uencing workflows. Graph-based data structures are suitable for compactly
representing multiple closely related reference genomes. Pangenome graphs
use a directed graph format\, where vertices are labelled with strings\, a
nd the individual reference genomes are represented as paths in the graph.
Aligning sequences (reads) to pangenome graphs is fundamental for pangeno
me-based genome resequencing.\n\nThe sequence-to-graph alignment problem s
eeks a walk in the graph that spells a sequence with minimum edit distance
from the input sequence. Unfortunately the exact algorithms known for sol
ving this problem are not scalable. Among the known heuristics\, co-linear
chaining is a common technique for quickly aligning reads to a graph. How
ever\, the known chaining algorithms are restricted to directed acyclic gr
aphs (DAGs) and are not trivially generalizable to cyclic graphs. Addressi
ng this limitation is important because pangenome graphs often contain cyc
les due to inversions\, duplications\, or copy number mutations within the
reference genomes.\n\nThis thesis presents the first practical formulatio
n and algorithm for co-linear chaining on cyclic pangenome graphs. Our wor
k builds upon the known chaining algorithms for DAGs. We propose a novel i
terative algorithm to handle cycles and provide a rigorous proof of correc
tness and runtime complexity. We also use the domain-specific small-width
property of pangenome graphs. The proposed optimizations enable our algori
thm to scale to large human pangenome graphs. We implemented the algorithm
in C++ and referred to it as PanAligner (https://github.com/at-cg/PanAlig
ner). PanAligner is an end-to-end long-read aligner for pangenome graphs.
We evaluated its speed and accuracy by aligning simulated long reads to a
cyclic human pangenome graph comprising 95 haplotypes. We achieved superio
r read mapping accuracy compared to existing methods.\n\n\n\n\n\n\nALL ARE
WELCOME\n\n
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:60@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240705T145000
DTEND;TZID=Asia/Kolkata:20240705T155000
DTSTAMP:20240701T131408Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-online-mode-cds-05-j
uly-2024data-driven-approach-to-estimate-wcet-for-real-time-systems/
SUMMARY:Ph.D: Thesis Defense: ONLINE MODE: CDS: 05\, July 2024"Data-Driven
Approach to Estimate WCET for Real-Time Systems"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Def
ense\n\n\n\nSpeaker : Mr. Vikash Kumar\nS.R. Number : 06-18-02-10-12-18-1-
16344\nTitle : "Data-Driven Approach to Estimate WCET for Real-Time System
s"\nResearch Supervisor : S K Nandy (retd) and S Raha\, CDS\nDate &\; T
ime : July 05\, 2024 (Friday)\, 02:50 PM\nVenue : The Thesis Defense will
be held on MICROSOFT TEAMS\nPlease click on the following link to join the
Thesis Defense:\nMS Teams link\n\n\n\nABSTRACT\n\nEstimating Worst-Case E
xecution Time (WCET) is paramount for developing Real-Time and Embedded sy
stems. The operating system’s scheduler uses the estimated WCET to sched
ule each task of these systems before the assigned deadline\, and failure
may lead to catastrophic events such as resource damage or even life loss.
These systems must satisfy the timing constraints. For instance\, it is e
ssential to know that car airbags open fast enough to save lives. The majo
r components required to estimate WCET are architecture or platform\, appl
ication\, and worst-case data. In this regard\, we propose novel methods f
or these components using machine learning techniques to estimate WCET saf
ely and precisely to make these systems more predictable and reliable than
traditional approaches.\n\n• Estimation of WCET on GPU architecture: Wi
th the advances in machine learning and artificial intelligence in every f
ield of life\, due to its tendency to solve many problems with accuracy\,
it requires Graphics Processing Units (GPUs) to provide massive parallelis
m for computation. GPUs are designed to provide high-performance through-p
ut\, but their integration into real-time systems focuses on predictabilit
y because most safety-critical applications have strict deadlines that nee
d to be followed to avoid unwanted situations. We propose a Machine Learni
ng approach to estimate the WCET of the GPU kernel from the binary of the
applications. The approach helps reduce the significant design space explo
ration in a short time. We use a measurement-based approach to train the m
achine-learning model using different kernel instructions\, which can pred
ict the WCET of the GPU kernel to detect timing misconfiguration in the la
ter development phase of the systems.\n\n• Estimation of WCET on Mixed-C
riticality Systems: In Mixed-Criticality (MC) Systems\, there is a trend o
f having multiple functionalities upon a single shared computing platform
for better cost and power efficiency. In this regard\, estimating the suit
able optimistic WCET based on the different system modes is essential to p
rovide these functionalities. A single application has assigned multiple W
CETs based on the criticality of the system\, such as safety-critical\, mi
ssion-critical\, and non-critical. We propose ESOMICS\, a novel method to
estimate suitable optimistic WCET using a Machine Learning model. Our appr
oach is based on the application’s source\, and the model is trained bas
ed on the large data set. To prove the effectiveness of our approach\, we
evaluated it with a newly defined metric EDT using an analytical solution
that allows us to compute the optimum value in a mixed-criticality system
based on experimentation. Our experimental results outperform all the prev
ious state-of-the-art approaches.\n\n• Estimation of Worst-Case Data for
WCET: Worst-Case Data which gives maximum execution time\, plays a vital
role in the estimation of WCET. An evolutionary algorithm\, such as the Ge
netic Algorithm\, has been employed to generate the Worst-Case Data. The c
omplexity of an evolutionary algorithm requires the use of several computa
tional resources. We propose a method to replace the hardware and simulato
r used in the evolution process with Machine Learning models. This method
reduces the overall time required to generate Worst-Case Data. Different m
achine learning models are trained to integrate with genetic algorithms. T
he feasibility of the proposed approach is validated using benchmarks from
different domains. The results show the speedup in the generation of Wors
t-Case Data.\n\n• Estimation of Early WCET: WCET is available to us in t
he last stage of systems development when the hardware is available\, and
the application code is compiled. Different methodologies measure the WCET
\, but none give early insights into WCET\, whichis crucial for system dev
elopment. If the system designers overestimate WCET in the early stage\, t
hen it would lead to an overqualified system\, which will increase the cos
t of the final product\, and if they underestimate WCET in the early stage
\, then it would lead to financial loss as the system would not perform as
expected. We propose to estimate early WCET using Machine Learning and De
ep Neural Networks as an approximate predictor model for hardware architec
ture and compiler. This model predicts the WCET based on the source code w
ithout compiling and running on the hardware architecture. The resulting W
CET needs to be revised to be used as an\nupper bound on the WCET. However
\, getting these results in the early stages of system development is an e
ssential prerequisite for the system’s dimension’s and configuration o
f the hardware setup.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:59@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240627T100000
DTEND;TZID=Asia/Kolkata:20240627T110000
DTSTAMP:20240626T144018Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-102-cds-27-june-2
024-improving-the-efficiency-of-variational-pinns-and-its-applications-to-
fluid-flow-problems/
SUMMARY:Ph.D: Thesis Colloquium: 102 : CDS: 27\, June 2024 "Improving the E
fficiency of Variational PINNs and its applications to fluid flow problems
"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Co
lloquium\n\n\n\n\n\n\nSpeaker : Mr. Thivin Anandh D\nS.R. Nu
mber : 06-18-01-10-12-18-1-15722\nTitle : "Impr
oving the Efficiency of Variational PINNs and its applications to fluid fl
ow problems"\nResearch Supervisor : Prof. Sashikumaar Ganesan\nDate &\;
Time : June 27\, 2024 (Thursday)\, 10:00 AM\nVenue
: # 102 CDS Seminar Hall\n\n\n\n\n\n\n\nABSTRACT\n\nFastVPINNs: A Tens
or-Driven Accelerated framework for Variational Physics informed neural ne
tworks in complex domains: Variational Physics-Informed Neural Networks (V
PINNs) utilize a variational loss function to solve partial differential e
quations\, mirroring Finite Element Analysis techniques. Traditional hp-VP
INNs\, while effective for high-frequency problems\, are computationally i
ntensive and scale poorly with increasing element counts\, limiting their
use in complex geometries. This work introduces FastVPINNs\, a tensor-base
d advancement that significantly reduces computational overhead and handle
s complex geometries. Using optimized tensor operations\, FastVPINNs achie
ve a 100-fold reduction in the median training time per epoch compared to
traditional hp-VPINNs. With proper choice of hyperparameters\, FastVPINNs
can surpass conventional PINNs in speed and accuracy\, especially in probl
ems with high-frequency solutions. We have also demonstrated solving inver
se problems(constant parameter inverse and domain inverse) for scalar PDEs
.\n\nA Open-Source PyPI package for FastVPINNs: This work presents the imp
lementation details of the FastVPINNs library as a Python pip package. Dev
eloped using TensorFlow 2.0\, the package now supports 3D scalar problems\
, making it one of the first hp-VPINNs frameworks to support 3D problems o
n complex geometries. The library includes a comprehensive test suite with
unit\, integration\, and compatibility tests\, achieving over 96% code co
verage. It also features CI/CD actions on GitHub for streamlined deploymen
t. Documentation is available at https://cmgcds.github.io/fastvpinns.\n\nF
astVPINNs for Flow problems (Navier Stokes): The incompressible Navier-Sto
kes equations (NSE) are essential for solving fluid dynamics problems. Whi
le PINNs have been used to solve NSE problems\, there is no literature on
VPINNs due to challenges such as the need for a higher number of elements
for vector-valued problems and the complexity of implementing variational
PINNs for the three components of the equations. These issues also lead to
infeasible training times with existing implementations. In this work\, w
e implement NSE using FastVPINNs and compare our results with PINNs in ter
ms of accuracy and training time. We solve forward problems such as a lid-
driven cavity\, flow through a channel\, Falkner-Skan boundary layer\, flo
w past a cylinder\, flow past a backward-facing step\, and Kovasznay flow
for Reynolds numbers ranging from 1 to 200 in the laminar regime. Our expe
riments show that FastVPINNs code runs twice as fast as PINNs and achieves
accuracy comparable to results in the literature. Additionally\, we solve
inverse problems for the NSE\, identifying the Reynolds number of the flo
w based on sparse solution observations.\n\nFastVPINNs for Singularly-Pert
urbed problems: Singularly-perturbed problems arise in convection-dominate
d regimes and are challenging test cases to solve due to the spurious osci
llations that might occur while solving the problem with conventional nume
rical methods. Stabilization schemes like Streamline-Upwind Petrov-Galerki
n (SUPG) and cross-wind loss functionals enhance numerical stability. Sinc
e SUPG stabilization is proposed in the weak formulation of PDEs\, Variati
onal PINNs are a suitable candidate for solving these problems. In this wo
rk\, we explore different stabilization schemes and their effects on singu
larly-perturbed problems\, comparing the accuracy of our results with the
existing literature. We demonstrate that stabilized VPINNs perform better
than PINNs proposed in the literature. Additionally\, we propose an neural
network model that predicts the SUPG stabilization parameter along with t
he solution\, addressing a challenging task in conventional methods. We al
so explore adaptive hard constraint functions for boundary layer problems\
, using neural networks to adjust the slope based on diffusion coefficient
s\, improving accuracy and reducing the need for tuning hyperparameters.\n
\nDomain-decomposition-based distributed training approach for FastVPINNs:
Variational Physics-Informed Neural Networks (VPINNs) can be computationa
lly expensive to train\, especially on larger domains with many elements.
To address this\, a domain-decomposition based training approach\, known a
s Finite Basis PINNs\, was proposed in the literature. We extend this appr
oach to Variational PINNs with domain decomposition. In FBVPINNs(Finite Ba
sis VPINNs)\, the domain is divided into subdomains and each subdomain is
assigned to a separate neural network\, with information exchange between
subdomains managed by aggregating gradients and solutions in overlapping r
egions using smooth\, differentiable window functions. This approach trans
forms complex global optimization into smaller optimization problems\, sig
nificantly reducing training times and addressing spectral bias on higher
frequency problems. Additionally\, we present an MPI-based implementation
of FBVPINNs for distributed training for lower frequency solution problems
.\n\n\n\n\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:57@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240621T110000
DTEND;TZID=Asia/Kolkata:20240621T120000
DTSTAMP:20240610T180046Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-online-mode-cds-21-j
une-2024modeling-physiological-transport-at-scales-connecting-cells-to-org
ans/
SUMMARY:Ph.D: Thesis Defense: ONLINE MODE: CDS: 21\, June 2024"Modeling phy
siological transport at scales: connecting cells to organs"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Def
ense\n\n\n\nSpeaker : Ms. Deepa Maheshvare M\nS.R. Number : 06-18-01-10-12
-16-1-14025\nTitle : "Modeling physiological transport at scales: connecti
ng cells to organs"\nResearch Supervisor : Prof. Debnath Pal\nDate &\;
Time : June 21\, 2024 (Friday)\, 11:00 AM\nVenue : The Thesis Defense will
be held on MICROSOFT TEAMS\nPlease click on the following link to join th
e Thesis Defense:\nMS Teams link\n\n\n\nABSTRACT\nThe physiological system
is a complex network in which each organ forms a subsystem\, and the func
tional networks in different subsystems communicate to maintain body's ove
rall homeostasis. The ability to simultaneously capture local and global d
ynamics by hierarchically bridging communication networks at different sca
les is a key challenge in holistic physiology modeling.\nWe present a scal
able hierarchical framework that allows us to bridge diverse scales to mod
el biochemicals' production\, consumption\, and distribution in tissue mic
roenvironments. We developed a discrete modeling framework to simulate the
gradient-driven advection–dispersion-reaction physics of multispecies t
ransport in multiscale systems. The physical space is translated into a me
tamodel\, and we define graph operators on the finite connected network re
presentation of the discrete functional units embedded in the metamodel. T
he governing differential equations capture the inter-compartment dynamics
of the well-mixed nodal volumes by formulating the transport dynamics in
the vascular domain\, transcapillary exchange\, and metabolism in the tiss
ue domain as a 'tank-in-series' model. This allows our framework to scale
to large networks and provides the flexibility to fuse multiscale models b
y encoding imaging data of vascular topology and omics data to enhance sys
tems-level understanding. Our framework is suitable for reducing the compu
tational cost of spatially discretizing large tissue volumes and for probi
ng the effect of flow topology on biochemical transport to study structure
-function relationships in tissues.\nNext\, we developed a comprehensive a
nd standardized data-driven modeling workflow to address the challenges fa
ced in developing kinetic models of metabolism in single cells. We have cr
eated open\, free\, and FAIR (findable\, accessible\, interoperable\, and
reusable) assets to study pancreatic physiology and glucose-stimulated ins
ulin secretion (GSIS). The data curation\, integration\, normalization and
data fitting workflow\, and a large database of metabolic data from 39 st
udies spanning 50 years of pancreatic\, islet\, and β-cell research in hu
mans\, rats\, mice\, and cell lines were used to construct a novel data-dr
iven kinetic SBML (Systems Biology Markup Language) model. The model consi
sts of detailed glycolysis and phenomenological equations for biphasic ins
ulin secretion coupled to ATP dynamics\, and (ATP/ADP ratio). The predicti
ons of glycolytic intermediates and biphasic insulin secretion are in good
agreement with experimental data\, and our model predicts the factors aff
ecting ATP consumption\, ATP formation\, hexokinase\, phosphofructokinase\
, and ATP/ADP-dependent insulin secretion influence GSIS.\nFinally\, we pr
esent KiPhyNet\, an online network simulation tool connecting cellular kin
etics and physiological transport. It allows users to simulate and interac
tively visualize pressure\, velocity\, and concentration fields for applic
ations such as flow distribution\, glucose transport\, and glucose-lactate
exchange in microvascular networks. When extended for translational purpo
ses in clinical settings\, the framework and pipeline developed in this wo
rk can advance the simulation of whole-body models and are expected to hav
e major applications in personalized medicine.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:55@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240618T103000
DTEND;TZID=Asia/Kolkata:20240618T113000
DTSTAMP:20240610T175649Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-cds-application-s
ervice-resilience-framework-an-end-to-end-perspective/
SUMMARY:Ph.D. Thesis {Colloquium}: CDS: "Application Service Resilience Fra
mework: An end-to-end perspective."
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Co
lloquium \n\n\n\n\n\nSpeaker : Ms. Dhanya R Math
ews\nS.R. Number : 06-18-02-10-12-18-1-15855\nTitle
: "Application Service Resilience Framework: An e
nd-to-end perspective "\nResearch Supervisor: Dr. J. Lakshmi\nDate &
\; Time : June 18\, 2024 (Tuesday) at 10:30 AM\nVenue
: The Thesis Colloquium will be held on HYBRID
Mode\n # 102 CDS Seminar
Hall /MICROSOFT TEAMS.\nPlease click on the following link to join the T
hesis Colloquium:\nMS Teams link\n\n\n\n\n\n\n\n\n\n\nABSTRACT\n\n\nThe id
ea of computing as a utility was realized with the emergence of the cloud
computing paradigm. Cloud service providers offer a wide range of services
that are delivered over the Internet to cloud service consumers. In its c
urrent manifestation\, the Cloud services are realized over multiple logic
al\, virtualized\, and distributed resources\, typically using a multi-lay
ered architecture. The providers document the non-functional service level
guarantees like availability\, performance\, security\, etc\, in Service
Level Agreements (SLAs) provided to the consumer as Service Level Objectiv
es (SLO). The wide adoption of cloud computing\, compounded with the emerg
ence of microservice architecture\, has resulted in a considerable increas
e in the number of components involved in service delivery. Manually addre
ssing failures in real-time is inefficient and often impossible at the clo
ud scale\, where failures are a norm rather than an exception. Ensuring th
e quality of an application service\, as documented in the SLA\, therefore
requires autonomous mechanisms to enhance cloud services' resilience.\n\n
Though cloud setups rely on highly autonomous service layers for managing\
, provisioning\, and monitoring applications\, most of them focus on a spe
cific cloud service architecture layer or consider only a particular set o
f faults. Any component across the cloud service stack involved in the ser
vice delivery could disrupt the SLO. Further\, as cloud services use share
d infrastructure\, monitoring and acting on the individual service layer m
etrics is limiting. In such a scenario\, the visibility of failure anywher
e in the stack can offer effective recovery/remediation strategies\; hence
\, an application-oriented approach that takes an end-to-end view of failu
res makes the case for any resiliency solution. Towards this\, we propose
an end-to-end service resilience framework that employs data-dependent int
elligent autonomous mechanisms to deal with cloud service disruptions effi
ciently. The intelligence to reduce the effect of disruptions is based on
understanding the complex interconnections and inter-dependencies of end-t
o-end components in the cloud service stack.\n\nThe different cloud servic
e abstraction layers and infrastructure sharing have resulted in increased
occurrence of faults\, more specifically\, saturation faults. The initial
phase of this work examines real-world disruption scenarios to understand
the faults that could disrupt a cloud service. With ever-changing applica
tions and environments on which they are hosted\, realizing a failure repo
sitory for cloud service faults is infeasible. This makes conventional dat
a-oriented approaches less practical and dynamic observability data-orient
ed methods more desirable. Towards this\, the second phase of this work de
veloped a Topology Aware Root Cause Detection Algorithm (TA-RCD) that cons
iders the observability data from end-to-end service components and their
interconnectedness. Our results from the fault injection studies show that
the proposed approach performs better than the state-of-the-art RCD algor
ithm\, at least by 2x times for Top-5 recall and 4x times for Top-3 recall
\, on average.\n\nTo autonomously recover a service from its anomalous sta
te\, the remediation should target the root cause of anomalous behavior. T
he root-cause localizations\, though accurate\, are not restricted to a sp
ecific component because of causal effects due to service interactions. In
order to identify the anomalous component\, the third phase of this work
developed a Topology Aware end-to-end failure Recovery framework (TA-REC)
that identifies the appropriate remediation strategy for an anomaly. The a
nomaly scores assignment and component activity tracking in TA-REC facilit
ates the identification of the component and the remediation that needs to
be applied to the component. For the saturation fault scenarios injected
across the stack\, TA-REC can identify an adequate remediation/recovery st
rategy than the state-of-the-art because of the better visibility of the o
rigin of the failure. The end-to-end visibility hence enables TA-REC to be
effective against an anomaly.\n\nIn conclusion\, this work demonstrated t
he usefulness of the end-to-end topology of a cloud application service to
remediate anomalies that challenge the service quality efficiently. The o
bservations prove that looking at the service as a black box restricts the
development of intelligent autonomous approaches to guarantee SLOs. The p
roof-of-concept evaluations demonstrated that the intelligence to maintain
service resilience effectively is based on an accurate understanding of t
he end-to-end state\, as it facilitates maintaining component serviceabili
ty by targeting the cause of failure in the stack. Future work aims to eva
luate both TA-RCD and TA-REC for a broader range of fault scenarios in rea
l-life production deployments.\n\n\n\n\n\n\nALL ARE WELCOME\n\n
CATEGORIES:Events,Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:56@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240611T100000
DTEND;TZID=Asia/Kolkata:20240611T110000
DTSTAMP:20240610T175831Z
URL:https://cds.iisc.ac.in/events/seminar-cds-202-11th-june-large-small-ed
dy-simulations-a-practical-method-for-high-fidelity-simulations-of-large-r
eynolds-number-turbulent-flows/
SUMMARY:{Seminar} @ CDS: #202: 11th June: "Large/Small Eddy Simulations: A
practical method for high-fidelity simulations of large Reynolds number tu
rbulent flows."
DESCRIPTION:Department of Computational and Data Sciences\nDepartment Semin
ar\n\n\n\nSpeaker : Arnab Moitro\, PhD Candidate\, University of Connectic
ut\, USA\nTitle : Large/Small Eddy Simulations: A practical method for hig
h-fidelity simulations of large Reynolds number turbulent flows.\nDate &am
p\; Time : June 11\, 2024\, 10:00 AM\nVenue : # 202\, CDS Classroom\n\n\n\
nABSTRACT\nTurbulent flows are ubiquitously found in nature and engineerin
g systems and are characterized b chaotic fluctuations spanning a wide r
ange of interacting scales. Numerical simulation of turbulent flows from f
irst principles using direct numerical simulations (DNS) is computationall
y expensive and restricted to moderate Reynolds numbers in idealized domai
ns\, decoupled from any realistic large-scale flow. On the other hand\, lo
wer fidelity techniques such as large eddy simulations (LES)\, used for mo
delling practical and natural systems rely on closure models that make mul
tiple assumptions which are often violated and require prior knowledge of
the flow. In this talk\, a novel multi-fidelity approach will be discussed
which couples a lower-fidelity\, unresolved\, time-dependent calculation
of the entire system (LES)\, and a high-fidelity\, fully resolved simulati
on of a sub-region of interest of the LES. The method is formulated in phy
sical space\, makes no assumptions of equilibrium\, isotropy\, initial or
boundary conditions of the underlying flow\, and can potentially be used f
or any regime whose properties are unknown. A priori and posteriori tests
of both steady and unsteady homogeneous\, isotropic turbulence are used to
demonstrate the method accuracy in recovering turbulence properties\, inc
luding spectra\, high order moments of velocity gradients\, and probabilit
y density functions of the intermittent quantities. The method is shown to
achieve DNS-level accuracy with three orders of magnitude reduction in co
mputational cost\, thus opening the possibility to study high Reynolds num
ber flow regimes.\n\nBIOGRAPHY\nArnab Moitro is a PhD candidate in the Sch
ool of Mechanical\, Aerospace and Manufacturing Engineering at the Univers
ity of Connecticut. Prior to this\, he received his dual degree (BTech+MTe
ch) from Indian Institute of Technology\, Madras and also worked as a rese
arch associate at Indian Institute of Science\, Bangalore. He will be join
ing Newcastle University as a post-doctoral research associate in July 202
4. His work focusses on high- and multi-fidelity simulations and modelling
of turbulent reacting and non-reacting flows.\n\nHost Faculty: Dr. Kondur
i Aditya\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:54@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240611T100000
DTEND;TZID=Asia/Kolkata:20240611T110000
DTSTAMP:20240604T081213Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-cds-designing-qua
lity-of-service-aware-serverless-platforms/
SUMMARY:Ph.D. Thesis {Colloquium}: CDS: "Designing Quality of Service aware
Serverless Platforms."
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\n\nPh.D. Thesis
Colloquium\n\n\n\n\nSpeaker : Mr. Sheshadri Kalk
unte Ramachandra\nS.R. Number : 06-18-02-10-12-18-1-16336\nTi
tle : "Designing Quality of Service aw
are Serverless Platforms "\nResearch Supervisor: Dr. J. Lakshmi\nDate &
amp\; Time : June 11\, 2024 (Tuesday) at 10:00 AM\nVenue
: The Thesis Colloquium will be held on HYBR
ID Mode\n # 102 CDS Sem
inar Hall /MICROSOFT TEAMS.\nPlease click on the following link to join th
e Thesis Colloquium:\nMS Teams link\n\n\n\n\nABSTRACT\n\nServerless Compu
ting is a highly abstract computing service that allows users to design th
eir applications as a workflow of independent stateless functions. The ser
verless platforms provide event-driven function executions and a powerful
pay-as-you-use billing model. The arduous task of function deployment\, re
source provisioning\, management\, and system administration is completely
undertaken by the serverless service provider\, thus providing ease of us
e to the user. Though serverless platforms show many benefits and are wide
ly adopted by many application domains\, they also exhibit certain shortco
mings. In a bid to provide clean programming abstractions and ease of use\
, the existing platforms suffer from significant platform overheads\, lead
ing to non-deterministic application execution latency. Shared resources\,
independent execution of individual workflow components\, interaction bet
ween various control components\, external data exchange\, etc. are some o
f the major contributors to the overheads. In addition\, the prevalence of
resource heterogeneity in serverless platforms\, together with the platfo
rm overheads\, gives rise to variable performance in serverless function e
xecutions. The varying application characteristics further add to the non-
determinism of the function executions. In this context\, Quality of Ser
vice (QoS) can be an important way of conveying application requirements f
or meeting its execution expectations. As QoS requirements are sensitive t
o each application domain\, to understand its impact\, an exploration of h
ow QoS requirements can be handled in serverless platforms is conducted in
this work. The study identifies the existing literature for selecting the
relevant application domains that are suitable for studying the impact of
QoS requirements on resource provisioning\, instance sizing\, function de
ployment decisions\, and data exchange among functions.\n\nIn the first ph
ase\, we examine the domain of Image and Video analytics. The applications
in this domain are typically deployed in a heterogeneous resources settin
g akin to the Edge-Cloud continuum. With varying application input size\,
it is imperative to select resource assignments to functions that meets us
er-specified QoS requirements. However\, in a heterogeneous resource pool
comprising Edge and Cloud resources\, choosing a resource layer and functi
on resource size that is both cost-efficient and one that meets QoS requir
ements is a challenging task. In order to address this issue\, we present
a novel QoS-aware serverless platform that deduces function resource speci
fication in a heterogeneous resource setting. The proposed platform perfor
ms function placement across Edge and Cloud layers based on incoming input
size and user-specified QoS requirements while prioritizing low-cost func
tion executions. Experimental results based on real-world workloads on a v
ideo surveillance application show that the proposed platform brings effic
ient resource utilization and cost savings at the Cloud by reducing the re
source usage by up to 30%.\n\nIn the second phase\, we examine the domain
of Network Function Virtualization. Network Function Virtualization (NFV
) and Software Defined Networks (SDNs) allow the network functions to be p
rogrammed as software entities that can be deployed on commodity servers i
n the Cloud\, referred to as Virtual Network Functions (VNFs). VNFs are ar
ranged in a workflow to enforce a network policy referred to as Service Fu
nction Chain (SFC). Existing solutions examine the SFC deployment on eithe
r IaaS or FaaS offerings. IaaS solutions alone lack the dynamism to respon
d to quick workload changes for network traffic which is similar to stream
ing data. Existing FaaS solutions exhibit significant platform overheads w
hich are prohibitive to meet the application's requirements. Further\, the
variability in the incoming flowlet and payload sizes of applications exa
cerbates the problem of selecting function resource assignments that both
satisfy QoS requirements and one that results in low resource usage. In or
der to address these shortcomings\, we propose a Hybrid Serverless Platfor
m (HSP) encompassing IaaS and FaaS in its function deployment strategy. Sp
ecifically\, The IaaS components handle the steady state load\, whereas th
e FaaS components activate during the dynamic change associated with scali
ng to minimize service loss. The proposed HSP controller takes resource pr
ovisioning decisions based on user-specified QoS requirements and the appl
ication's flow statistics. The HSP controller's design exploits data local
ity in SFC realization\, reducing data-transfer times and platform overhea
ds between VNFs. The proposed platform provides up to 35% resource savings
as compared to a pure IaaS deployment and up to 55% lower end-to-end SFC
execution times as compared to a baseline FaaS implementation with minimal
loss of flowlet service.\n\nIn the third phase\, we examine the domain of
Web Inferencing which constitutes an important class of event-driven appl
ications that lend themselves to being composed as a workflow of functions
. Typically\, web inferencing applications are deployed across a heterogen
eous setting comprising Edge and Cloud resources. Though web inferencing w
orkflows have seen an extensive adoption of serverless computing\, they su
ffer from the overheads of the serverless computing platforms\, which incr
eases the application's end-to-end latency. Existing literature addresses
the problem of high platform overheads by a technique called 'Function Fus
ion'\, which combines consecutive serverless functions in a workflow into
a single logical group. However\, existing solutions are limited by the pe
ak resource sizes of functions that dictate the sizing of fusion groups\;
they are insensitive to the varying application inputs and QoS requirement
s of the inferencing applications. In order to address these issues\, we p
ropose a novel fusion criteria called 'RightFusion' that performs cost-eff
icient function fusion in serverless platforms that manage a heterogeneous
resource pool. The proposed platform uses incoming application characteri
stics and user-specified QoS requirements to dynamically arrive at functio
n fusion decisions that are both cost-efficient and QoS-abiding. Experimen
tal evaluation of representative inferencing applications shows that Right
Fusion reduces resource usage by above 35% at Edge and Cloud while meeting
the QoS requirements compared to the PeakFusion baseline.\n\nIn conclusio
n\, this study demonstrates the value of considering application QoS in se
rverless computing. The QoS specifications enable the platform to undertak
e resource management and function placement decisions that meet user expe
ctations. In our studies\, QoS requirements from different application dom
ains have translated to different design decisions that address various pr
oblems\, including resource right-sizing\, function fusion\, request order
ing\, etc.\, resulting in cost and resource savings. Current Serverless pl
atforms take off the burden of resource provisioning and placement for exe
cution from the user. However\, they are agnostic to the expectations of a
pplications. Applications\, on the other hand\, suffer from non-determinis
tic overheads from these platforms. Given the differences in differing app
lications\, QoS provides a mechanism to express application requirements t
hat the serverless platform can use for decision-making with regard to app
lication execution. The current exploration exposes that QoS is applicatio
n-specific and requires different methods to deal with it\; emphasizing th
e need for customization in serverless platform design. It also brings to
the fore the value of considering the integration of learning approaches i
n serverless controller's decision-making. This study considered cost and
performance as the primary QoS attributes for each of the domains examined
. Other QoS features\, like security\, privacy\, reliability\, jurispruden
ce\, etc.\, are yet to be explored.\n\n\n\n\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:43@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata;VALUE=DATE:20240531
DTEND;TZID=Asia/Kolkata;VALUE=DATE:20240602
DTSTAMP:20240323T082557Z
URL:https://cds.iisc.ac.in/events/symposium-on-big-data-algorithms-for-bio
logy-bdbio/
SUMMARY:Symposium on Big Data Algorithms for Biology (BDBio)
DESCRIPTION:Join us at the second annual Symposium on Big Data Algorithms f
or Biology (BDBio) on May 31-June 1\, 2024 at the Faculty Hall\, Main Buil
ding\, Indian Institute of Science.\n\nThe 2-day event will feature engagi
ng invited talks\, contributed talks\, poster presentations\, and panel di
scussions on various topics in computational biology. Registration / Abstr
act submission are open until April 20. Check more details on www.bdbio.in
\n\n
CATEGORIES:Events
END:VEVENT
BEGIN:VEVENT
UID:52@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240527T140000
DTEND;TZID=Asia/Kolkata:20240527T150000
DTSTAMP:20240523T070641Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-27th-may-genaudit-fixing
-factual-errors-in-language-model-outputs-with-evidence/
SUMMARY:{Seminar} @ CDS: #102: 27th May: "GenAudit: Fixing Factual Errors i
n Language Model Outputs with Evidence."
DESCRIPTION:Department of Computational and Data Sciences\nDepartment Semin
ar\n\n\n\nSpeaker : Kundan Krishna\, PhD candidate in the Language Technol
ogies Institute at CMU\nTitle : "GenAudit: Fixing Factual Errors in Langua
ge Model Outputs with Evidence."\nDate &\; Time : May 27\, 2024\, 02:00
PM\nVenue : # 102\, CDS Seminar Hall\n\n\n\nABSTRACT\nLLMs can generate f
actually incorrect statements even when provided access to reference docum
ents. Such errors can be dangerous in high-stakes applications (e.g.\, doc
ument-grounded QA for healthcare or finance). In this talk\, I would prese
nt GenAudit --- a tool intended to assist fact-checking LLM responses for
document-grounded tasks. GenAudit suggests edits to the LLM response by re
vising or removing claims that are not supported by the reference document
\, and also presents evidence from the reference for facts that do appear
to have support. To power GenAudit\, we trained models on a scpecially cre
ated dataset with high-quality human annotations. Comprehensive evaluation
by human raters shows that GenAudit can detect errors in 8 different LLM
outputs when summarizing documents from diverse domains\, outperforming GP
T-4 while also being much cheaper. GenAudit is available for public use at
https://genaudit.org\n\nBIOGRAPHY\nKundan Krishna is a PhD candidate in t
he Language Technologies Institute at CMU\, advised by Professor Zachary L
ipton and Professor Jeffrey Bigham. His research focuses on mitigating saf
ety issues in deployment of language models by improving their factual acc
uracy and reducing reliance on web-scale scraped data for their pretrainin
g. He has also worked extensively on various aspects of text summarization
\, including improving robustness to noise\, producing structured summarie
s\, generating topic-focused summaries etc. Prior to CMU\, he graduated wi
th a Bachelor's degree from IIT Kanpur\, and worked as a research engineer
at Adobe Research\n\nHost Faculty: Dr. Danish Pruthi\n\n\n\nALL ARE WELCO
ME
CATEGORIES:Events
END:VEVENT
BEGIN:VEVENT
UID:53@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240527T110000
DTEND;TZID=Asia/Kolkata:20240527T120000
DTSTAMP:20240523T071016Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-27th-may-high-coverage-i
nformation-extraction/
SUMMARY:{Seminar} @ CDS: #102: 27th May: "High-coverage information extract
ion."
DESCRIPTION:Department of Computational and Data Sciences\nDepartment Semin
ar\n\n\n\nSpeaker : Sneha Singhania\, PhD student at the Max Planck Instit
ute for Informatics and Saarland University in Germany\nTitle : "High-cove
rage information extraction."\nDate &\; Time : May 27\, 2024\, 11:00 AM
\nVenue : # 102\, CDS Seminar Hall\n\n\n\nABSTRACT\nStructured knowledge\,
in the form of entities and relations\, is a powerful asset for search\,
recommendations\, and data integration and is extensively used by differen
t stakeholders. However\, converting noisy internet content into crisp kno
wledge structures requires heavy-duty processing of vast amounts of data.
Using language models (LMs) for information extraction (IE) is mature but
still struggles to achieve both high precision and high recall\, limiting
their reliable usage. In my talk\, I will present three lines of work for
high-coverage IE from various knowledge sources. Firstly\, I will detail h
ow one can identify and filter content-rich web documents\, laying out our
approach to rank documents to automatically build knowledge bases (KBs).
Secondly\, I will discuss the emerging role of LMs as KBs\, using various
probing techniques. Finally\, I will introduce the L3X framework\, which e
xtracts a long list of entities from long documents using retrieval-augmen
ted LMs. Together\, these techniques help us better handle the unknowns an
d construct complete structured knowledge.\n\nBIOGRAPHY\nSneha Singhania i
s a PhD student at the Max Planck Institute for Informatics and Saarland U
niversity in Germany\, advised by Gerhard Weikum and Simon Razniewski. Her
research aims to close the knowledge gap between data sources and models
to generate reliable output. In the past\, she graduated from IIIT-Bangalo
re with a dual degree in Computer Science\, worked as a researcher at Acce
nture Labs\, and interned at Apple Research in Cupertino.\n\nHost Faculty:
Dr. Danish Pruthi\n\n\n\nALL ARE WELCOME
CATEGORIES:Events
END:VEVENT
BEGIN:VEVENT
UID:51@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240425T100000
DTEND;TZID=Asia/Kolkata:20240425T110000
DTSTAMP:20240417T072125Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-cds-seminar-hall-
102-learning-deep-neural-networks-from-limited-and-imperfect-data/
SUMMARY:Ph.D. Thesis {Colloquium}: CDS Seminar Hall # 102 "Learning Deep Ne
ural Networks From Limited and Imperfect Data"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Col
loquium\n\n\n\nSpeaker : Mr. Harsh Rangwani\n\nS.R. Number : 06-18-01-10-1
2-19-1-17477\n\nTitle :"Learning Deep Neural Networks From Limited and Imp
erfect Data"\nResearch Supervisor: Prof. Venkatesh Babu\nDate &\; Time
: April 25\, 2024 (Thursday) at 10:00 AM\nVenue : # 102 CDS Seminar Hall\n
\n\n\nABSTRACT\n\nDeep Neural Networks have demonstrated orders of magnitu
de improvement in capabilities over the years after AlexNet won the ImageN
et challenge in 2012. One of the major reasons for this success is the ava
ilability of large-scale\, well-curated datasets. These datasets (e.g.\, I
mageNet\, MSCOCO\, etc.) are often manually balanced across categories (cl
asses) to facilitate learning of all the categories. This curation process
is often expensive and requires throwing away precious annotated data to
balance the frequency across classes. This is because the distribution of
data in the world (e.g.\, internet\, etc.) significantly differs from the
well-curated datasets and is often over-populated with samples from common
categories. The algorithms designed for well- curated datasets perform su
boptimally when used to learn from imperfect datasets with long-tailed imb
alances and distribution shifts. For deep models to be widely used\, getti
ng away with the costly curation process by developing robust algorithms t
hat can learn from real-world data distribution is necessary. Toward this
goal\, we develop practical algorithms for Deep Neural Networks that can l
earn from limited and imperfect data present in the real world. This thesi
s is divided into four segments\, each covering a scenario of learning fro
m limited or imperfect data. The first part of the thesis focuses on Learn
ing Generative Models for Long-Tail Data\, where we mitigate the mode-coll
apse for tail (minority) classes and enable diverse aesthetic image genera
tions as head (majority) classes. In the second part\, we enable effective
generalization on tail classes through Inductive Regularization schemes\,
which allow tail classes to generalize as the head classes without enforc
ing explicit generation of images. In the third part\, we develop algorith
ms for Optimizing Relevant Metrics compared to the average accuracy for le
arning from long-tailed data with limited annotation (semi-supervised)\, f
ollowed by the fourth part\, which focuses on the effective domain adaptat
ion of the model to various domains with zero to very few labeled samples.
\n\nGenerative Models for Long-Tail Data. We first evaluate generative mod
els’ performance\, specifically variants of Generative Adversarial Netwo
rks (GANs) on long-tailed datasets. The GAN variants suffer from either mo
de-collapse or miss-class modes during generation. To mitigate this\, we p
ropose Class Balancing GAN with a Classifier in the Loop\, which uses a cl
assifier to asses the modes in generated images and regularizes GAN to pro
duce all classes equally. To alleviate the dependence on the classifier\,
following our observation that spectral norm explosion of Batch Norm param
eters is the major reason for mode collapse. We develop an inexpensive gro
up Spectral Regularizer (gSR) to mitigate the spectral collapse\, signific
antly improving the SotA conditional GANs (SNGAN and BigGAN) performance o
n long-tailed data. However\, we observed that class confusion was present
in the generated images due to norm regularization. In our latest work\,
NoisyTwins\, we factor the latent space as distinct Gaussian by design for
each class\, enforcing class consistency and intra-class diversity using
a contrastive approach (BarlowTwins). This helps to scale high-resolution
StyleGANs for ≥ 1000 class long-tailed datasets of ImageNet-LT and iNatu
ralist2019\, achieving state-of-the-art (SotA) performance.\n\nInducting R
egularization Schemes for Long-Tailed Data. While Data Generation is excit
ing for improving classification models on tail classes\, it often comes w
ith the cost of training an auxiliary GAN model. Hence\, a lightweight tec
hnique like enhancing loss weights (re-weighting) for tail classes while t
raining CNNs is practical to improve minority class performance. However\,
despite this\, the model only attains minima for the head class loss and
converges to saddle point for tail classes. We show that inducing inductiv
e bias of escaping saddles and converging to minima for tail classes\, usi
ng Sharpness Aware Minimization (SAM) significantly improves performance o
n tail classes. Further training Vision Transformer (ViT) for long-tail re
cognition is hard\, as they don’t have inductive biases like locality of
features\, which makes them data hungry. We propose DeiT-LT\, which intro
duces OOD and low-rank distillation from CNN to induce CNN-like robustness
into scalable ViTs for robust performance.\n\nSemi-Supervised Long-Tailed
Learning. The above methods work in supervised long-tail learning\, where
they avoid throwing off the annotated data. However\, the real benefit of
long-tailed methods could be leveraged when they utilize the extensive un
labeled data present (i.e.\, semi-supervised setting). For this\, we intro
duce a paradigm where we measure the performance using relevant metrics li
ke worst-case recall and recall H-mean on a held-out set\, and we use thei
r feedback to learn in a semi-supervised long-tailed setting. We introduce
Cost-Sensitive Self Training (CSST) generalizes self-training (e.g.\, Fix
Match\, etc.) based semi-supervised learning to long-tail settings with st
rong guarantees and empirical performance. The general trend these days is
to use self-supervised pre-training to obtain a robust model and then fin
e-tune it. In this setup\, we introduce SelMix\, an inexpensive fine-tunin
g technique to optimize the relevant metrics using pre-trained models. In
SelMix\, we relax the assumption that unlabeled distribution is similar to
the labeled\, making models robust to distribution shifts.\n\nEfficient D
omain Adaptation. The long-tail learning algorithms focus on limited data
setup and improving in-distribution generalization. Still\, for practical
usage\, the model must learn from imperfect data and perform well across v
arious domains. Toward this goal\, we develop Submodular Subset Selection
for Adversarial Domain Adaptation\, which carefully selects a few samples
to be labeled for maximally improving model performance in the target doma
in. To further improve the efficiency of the Adaptation procedure\, we int
roduce Smooth Domain Adversarial Training (SDAT)\, which converges to gene
ralizable smooth minima. The smooth minimum enables efficient and effectiv
e model adaptation across domains and tasks.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:49@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240418T150000
DTEND;TZID=Asia/Kolkata:20240418T160000
DTSTAMP:20240410T073116Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-defense-cds-18-ap
ril-2024-an-importance-sampling-in-n-sphere-monte-carlo-and-its-performanc
e-analysis-for-high-dimensional-integration/
SUMMARY:M.Tech Research: Thesis Defense: CDS: 18\, April 2024 "An importanc
e sampling in N-Sphere Monte Carlo and its performance analysis for high d
imensional integration"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research
Oral Examination\n\n\n\nSpeaker : Mr. Jawlekar Abhijeet Rajendra\n\nS.R. N
umber : 06-18-01-10-22-21-1-20128\n\nTitle : "An importance sampling in N-
Sphere Monte Carlo and its performance analysis for high dimensional integ
ration"\n\nThesis examiner : Prof. Manohar C.\, Dept. of Civil Engine
ering\, Indian Institute of Science.\nResearch Supervisor: Prof. Murugesan
Venkatapathi\nDate &\; Time : April 18\, 2024 (Thursday) at 03:00 PM\n
Venue : # 102 CDS Seminar Hall\n\n\n\nABSTRACT\n\nStatistical methods for
estimating integrals are indispensable when the number of dimensions (para
meters) become greater than ~ 10\, where numerical methods are unviable in
general. Well-known statistical methods like Quasi-Monte Carlo converge q
uickly only for problems with a small number of effective dimensions\, and
Markov Chain Monte Carlo (MCMC) methods incur a sharply increasing comput
ing effort with the number of dimensions 'n' that is bounded as O(n^5). Th
is bound on the dimensional scaling of computing effort in multiphase MCMC
is limited to domains with a given convex shape (determined by the limits
of integration). Note that the non-convexity and roughness of the boundar
ies of the domain are factors that adversely affect the convergence of suc
h methods based on a random walk in high dimensions.\n\nA different approa
ch to high-D integration using (1D) line integrals along random directions
coupled with a less-known volume transformation was suggested here at the
Institute by Arun et. al. This method called as N-sphere Monte Carlo (NSM
C) is agnostic to the shape and roughness of the boundary for any given di
stribution of extents of the domain from a reference point. While the dime
nsional scaling of computing in NSMC integration can be bound as O(n^3) fo
r any distribution of relative extents (and not a particular convex shape)
\, a similar bound does not exist for MCMC as any given extent distributio
n can represent numerous geometries where it may not converge. It was show
n earlier that when restricted to convex shapes where the extent density f
unctions become increasingly heavy tailed as 'n' increases\, the naive NSM
C may be more efficient than the multiphase MCMC only when n <\; ~100. T
his thesis has three contributions. 1) It is analytically shown that\, unl
ike MCMC\, the convergence of NSMC in the estimation of n-volume of a doma
in is not a necessary condition for its convergence in any other integrati
on over that domain. 2) A direct numerical comparison of the naive NSMC an
d the multiphase MCMC was performed for estimating n-volumes and different
types of integrands\, establishing this advantage in integration even ove
r typical convex domains when n <\; ~ 100. 3) A method for importance sa
mpling is suggested for NSMC with a demonstration of the improved performa
nce in higher dimensions for domains with heavy tailed extent density func
tions. In identifying and ensuring a local volume of interest is sampled a
dequately\, this method employs efficient sampling in high-D cones with a
target distribution.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:50@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240415T090000
DTEND;TZID=Asia/Kolkata:20240415T100000
DTSTAMP:20240412T113508Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-colloquium-cds-se
quence-alignment-to-cyclic-pangenome-graphs/
SUMMARY:M.Tech Research Thesis {Colloquium}: CDS: "Sequence Alignment to Cy
clic Pangenome Graphs"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research
Thesis Colloquium\n\n\n\nSpeaker : Ms. Jyotshna Rajput\n\nS.R. Number : 06
-18-01-10-22-21-1-19943\n\nTitle :"Sequence Alignment to Cyclic Pangenome
Graphs"\nResearch Supervisor: Dr. Chirag Jain\nDate &\; Time : April 15
\, 2024 (Monday) at 09:00 AM\n\nVenue : The Thesis Colloquium will be held
on MICROSOFT TEAMS\n\n \;\n\n\n\nABSTRACT\n\nThe growing availability
of genome sequences of several species\, including humans\, has created t
he opportunity to utilize multiple reference genomes for bioinformatics an
alyses and improve the accuracy of genome resequencing workflows. Graph-ba
sed data structures are suitable for compactly representing multiple close
ly related reference genomes. Pangenome graphs use a directed graph format
\, where vertices are labeled with strings\, and the individual reference
genomes are represented as paths in the graph. Aligning sequences (reads)
to pangenome graphs is fundamental for pangenome-based genome resequencing
. The sequence-to-graph alignment problem seeks a walk in the graph that s
pells a sequence with minimum edit distance from the input sequence. Howev
er\, exact algorithms for solving this problem are unlikely to scale due t
o the known hardness of this problem. Co-linear chaining is a well-studied
and commonly used heuristic alternative for quickly aligning reads to a g
raph. However\, the known chaining algorithms are restricted to directed a
cyclic graphs (DAGs) and are not trivially generalizable to cyclic graphs.
This limitation must be addressed because pangenome graphs often contain
cycles due to inversions\, duplications\, or copy number mutations within
the reference genomes.\n\nThis thesis presents the first practical formula
tion and algorithm for co-linear chaining on cyclic pangenome graphs. Our
algorithm builds upon the known chaining algorithms for DAGs. We propose a
novel iterative algorithm to handle cycles and provide rigorous proof of
correctness and runtime complexity. Our algorithm also exploits the domain
-specific small-width property of pangenome graphs. The proposed optimizat
ions enable our algorithm to scale to large human pangenome graphs. We imp
lemented our algorithm in C++ and referred to it as PanAligner (https://gi
thub.com/at-cg/PanAligner). PanAligner is an end-to-end long-read aligner
for pangenome graphs. We evaluated its speed and accuracy by aligning simu
lated long reads to a cyclic human pangenome graph comprising 95 haplotype
s. We achieved superior read mapping accuracy using PanAligner compared to
existing methods.\n\n\n\nALL ARE WELCOME
CATEGORIES:MTech Research Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:47@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240404T160000
DTEND;TZID=Asia/Kolkata:20240404T170000
DTSTAMP:20240327T102422Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-defense-cds-04-ap
ril-2024-intelligent-methods-for-cloud-workload-orchestration/
SUMMARY:M.Tech Research: Thesis Defense: CDS: 04\, April 2024 "Intelligent
Methods for Cloud Workload Orchestration"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\nM.Tech Researc
h Thesis Defense\n\n\n\nSpeaker : Mr. Prathamesh Saraf Vinayak\n\nS.R. Num
ber : 06-18-01-10-22-21-1-19717\nTitle : "Intelligent Methods for Cloud Wo
rkload Orchestration."\nResearch Supervisor :Dr. J. Lakshmi\nDate &\; T
ime : April 04\, 2024 (Thursday)\, 04:00 PM\n\nVenue : The Thesis Défense
will be held on MICROSOFT TEAMS\n\nPlease click on the following link to
join the Thesis Defense:\n\nMS Teams link\n\n\n\nABSTRACT\n\nCloud workloa
d orchestration is pivotal in optimizing the performance\, resource utiliz
ation\, and cost-effectiveness of applications in data centers. As modern
businesses and IT operations are migrating their businesses to the cloud\,
understanding the dynamics of cloud data centers has become indispensable
. Often\, two perspectives play a pivotal role in workload orchestration i
n data centers. One is from the cloud provider side\, whose goal is to pro
vision as many applications as possible on the available resources\, bidin
g to SLA constraints and increasing return on investment. Others are from
the side of enterprises and individual customers\, often referred to as en
d users\, whose primary objective is to ensure application performance wit
h a reduced deployment cost. Containerization has gained popularity for de
ploying applications on public clouds\, where large enterprises manage num
erous applications through thousands of containers placed onto Virtual Mac
hines (VMs). While the need for cost-efficient placement in cloud data cen
ters is undeniable\, the complexities involved in achieving this goal cann
ot be understated. This problem is usually modeled as a multi-dimensional
Vector Bin-packing Problem (VBP). Solving VBP optimally is NP-hard and pra
ctical solutions requiring real-time decisions use heuristics. This work e
xplores the landscape of cloud data centers\, emphasizing the significance
of efficient bin packing in achieving optimal cost and resource utilizati
on. Traditional methods\, including heuristics and optimal algorithms\, fa
ce limitations in handling continuous request arrivals and the dynamic nat
ure of cloud workloads. Integer Linear Programming (ILP)\, which can provi
de optimal solutions for small problem sizes with tens of requests\, may t
ake minutes to hours to complete\, even at such scales. Moreover\, optimal
algorithms inherently demand perfect knowledge of all current and future
requests to be placed within the bins\, rendering them unsuitable for the
dynamic and often unpredictable online placement scenarios prevalent in cl
oud setups.\n\nTo address these challenges\, this work introduces a novel
approach to solving VBP through Reinforcement Learning (RL)\, trained on t
he historical container workload trace for an enterprise\, a.k.a CARL (Cos
t-optimized container placement using Adversarial Reinforcement Learning).
The proposed work evaluates the effectiveness of CARL in comparison to tr
aditional methods. CARL leverages historical container workload traces\, l
earning from a semi-optimal VBP solver while optimizing VM costs. The cont
ributions of this research extend beyond traditional methods\, providing i
nsights into the advantages and disadvantages of heuristics\, optimal algo
rithms\, and learning approaches. We trained and evaluated CARL on workloa
ds derived from realistic traces from Google Cloud and Alibaba for placing
10\,000 container requests onto over 8000 VMs. CARL is fast\, making plac
ement decisions for request sets with 124 containers per second within 65m
s onto 1000s of potential VMs. It is also efficient\, achieving up to 13.9
8% lower VM costs than baseline heuristics for larger traces. To push the
boundaries further\, we use the Mixture of Experts (MoE) strategy in CARL\
, wherein we use multiple experts who help CARL learn placement policies o
f various approaches combined. Including an MoE strategy enhances CARL's a
daptability to changes in workload distribution\, ensuring competitive per
formance in scenarios with skewed resource needs or inter-arrival times.\n
\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:48@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata;VALUE=DATE:20240404
DTEND;TZID=Asia/Kolkata;VALUE=DATE:20240406
DTSTAMP:20240402T184435Z
URL:https://cds.iisc.ac.in/events/15th-eecs-research-students-symposium/
SUMMARY:15th EECS Research Students Symposium
DESCRIPTION:15th EECS Research Students Symposium will be held on April 4t
h and 5th\, 2024. This annual event has served as an excellent forum for
interaction among IISc's students\, faculty members\, and other top resea
rchers from industry and academia. The symposium covers research topics ra
nging from theoretical computer science to power engineering.\n\nFor more
information on the schedule and speakers\, please visit https://eecs.iisc.
ac.in/EECS2024 or see the flyer below.\n\n[pdf-embedder url="https://cds.i
isc.ac.in/wp-content/uploads/EECS_Symp_2024_Flyer.pdf"].\n\n \;
CATEGORIES:Events
END:VEVENT
BEGIN:VEVENT
UID:44@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240321T113000
DTEND;TZID=Asia/Kolkata:20240321T123000
DTSTAMP:20240305T044838Z
URL:https://cds.iisc.ac.in/events/cds-research-seminar-development-of-ai-t
ools-in-liver-disease-and-transplantation-current-and-future-prospects/
SUMMARY:CDS Research Seminar: Development of AI tools in Liver Disease an
d Transplantation: Current and Future Prospects
DESCRIPTION:CDS Research Seminar\n\n\n\nTitle: Development of AI tools in
Liver Disease and Transplantation: Current and Future Prospects\nSpeaker:
Dr. Mamatha Bhat\, MD\, MSc\, PhD\nHepatologist &\; Co-Lead of Transpl
ant AI initiative (TAI)\, Ajmera Transplant Centre\nAssociate Professor\,
Division of Gastroenterology &\; Hepatology\nDirector\, Clinician-Scien
tist Training Program (CSTP)\, Department of Medicine\, University of Toro
nto\nPartnerships &\; Engagement Lead\, Temerty Centre for AI in Resear
ch &\; Education in Medicine (T-CAIREM)\nDate &\; Time: March 21\, 2
024 (Thursday) &\; 11:30 AM - 12:30 PM\nRoom No: CDS 102 (CDS Seminar H
all)\n\n\n\nAbstract:\n\nArtificial Intelligence (AI) tools have been incr
easingly applied to clinical questions in transplant medicine in recent ye
ars. As we continue to push the limits of transplantation\, with the older
age and complex comorbidities of candidates\, there are many challenges t
hroughout transplant medicine that must be better addressed. Various facto
rs affect liver transplant pathology and outcomes\, including sex\, ethnic
ity\, genetics\, BMI\, diabetes\, and immunosuppressive regimens. There ex
ist complex\, non-linear patterns in laboratory tests that must be conside
red in conjunction with the complex clinical variables to predict outcome.
Additionally\, electronic health record data\, imaging technologies\, his
tology\, and ‘omics data have continued to expand the types of data avai
lable. These complex data points\, their hidden patterns and interrelation
ships can be uniquely leveraged with the use of AI tools. Longitudinal cha
nges in these variables are also being examined to provide a continuous re
assessment of risk along the timeline. Applications of AI in transplant me
dicine include waitlist prioritization\, donor-recipient matching\, and sh
ort-term/long-term outcome prediction. In this talk\, I will go over these
considerations with respect to application of AI in transplant medicine.
I will additionally discuss integration of ‘omics data\, as well as pers
pectives regarding clinical deployment of AI tools.\n\nShort Biography:\n\
nDr. Mamatha Bhat is a Hepatologist and leads the Transplant AI initiative
at the University Health Network's Ajmera Transplant Centre\, and Associa
te Professor of Medicine at the University of Toronto. Dr. Bhat completed
her medical school and residency training at McGill University. She then c
ompleted a Transplant Hepatology fellowship at the Mayo Clinic in Rocheste
r\, Minnesota\, followed by a Canadian Institutes of Health Research (CIHR
) Fellowship for Health Professionals\, through which she completed a PhD.
\n\nThe goal of Dr. Bhat’s research program is to improve long-term outc
omes of liver transplantation through a precision medicine approach. Her p
rogram is unique in using tools of Artificial Intelligence with bioinforma
tics to personalize the care of liver transplant recipients based on an im
proved biological understanding of the liver and disease after transplant.
Her interdisciplinary program has been supported by the Canadian Institut
es for Health Research (CIHR)\, Canadian Donation and Transplant Research
Program (CDTRP)\, Natural Sciences and Engineering Research Council (NSERC
)\, Canadian Liver Foundation (CLF)\, among others. Dr. Bhat is Partnershi
p &\; Engagement Lead for the Temerty Centre for AI Research and Educat
ion in Medicine (T-CAIREM)\, Chair of the International Liver Transplant S
ociety Basic and Translational Science Research committee\; on the Executi
ve committee of the CDTRP\, and an Associate Editor for the American Journ
al of Transplantation. Dr. Bhat has also been the recipient of recognition
s such as the 2022 CIHR-INMD-CASL Early Career Researcher prize\, CASL Res
earch Excellence award and the 2021 American Society of Transplantation Ba
sic Science Career Development Award.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:46@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240311T100000
DTEND;TZID=Asia/Kolkata:20240311T110000
DTSTAMP:20240310T165024Z
URL:https://cds.iisc.ac.in/events/cds-seminar10211th-march-non-equilibrium
-plasma-and-its-use-in-combustion-a-modeling-perspec6ve/
SUMMARY:CDS Seminar@102:11th March: Non-Equilibrium Plasma and its Use in C
ombustion: A Modeling Perspec6ve
DESCRIPTION:Department of computational and Data Sciences\nDepartment semin
ar\n\n\n\nTitle: Non-Equilibrium Plasma and its Use in Combustion: A Mode
ling Perspec6ve\nSpeaker: Taaresh S. Taneja\, Department of Mechanical Eng
ineering\, University of Minnesota\, USA\nDate and time: 11 March 2024 (M
onday)\, 10:00 AM\nVenue: CDS 102\n\n\n\nAbstract:\n\n\nNon-equilibrium or
Low Temperature Plasma is a state of a gas which is characterized by a di
fference in the energies of the electrons and other heavy species of the
gas. Such a plasma typically constitutes of gas molecules\, a relatively
lower density of ions and rotational\, vibrational\, and electronic excite
d states\, neutral radicals\, along with free electrons. Non-equilibrium
plasmas can exist at various gas pressures ranging from 0.1 – 106 Pa an
d gas temperatures ranging from 100 – 10000 K. Across these wide range
of conditions\, the physical and chemical properties of the plasma can var
y substantially – which make them extremely useful in diverse technolog
ical areas such as semiconductor manufacturing\, water treatment\, medica
l equipment sterilization\, nanomaterial synthesis\, chemical reforming\,
combustion assistance\, etc. Between 104 - 106 Pa and 300 – 5000 K\, th
ese non-equilibrium plasma discharges can be used for assisting combustio
n. This assistance is provided majorly through two channels – gas heati
ng and chemical radical production. Both these channels can be used to ign
ite renewable and carbon-free fuels such as ammonia (NH3)\, which is very
difficult to burn\, and stabilize flames in challenging conditions such
as gas turbines\, scramjet\, and rocket combustors. Moreover\, the chemic
al pathways introduced by the non-equilibrium of the gas\, can also help t
o lower\, or completely prevent emissions from combustion\, such as unbur
ned hydrocarbons\, soot\, CO2\, CO and NOx. High fidelity computational s
imulations (DNS / LES) of non-equilibrium “plasma assisted combustion
” (PAC) face various constraints due to the wide-ranging temporal (10-15
– 10-2 s) and spatial (10-6 – 10-1 m) scales of this problem\, which
make the system of governing equations very stiff. Furthermore\, the hig
hly coupled interaction of electrostatics\, plasma chemistry\, combustion
chemistry\, and turbulent flow renders PAC its multi-physics nature. Thi
s talk will provide an overview of the governing physics\, the mathematic
al formulation of different models\, and a few technological applications
of non- equilibrium plasma assisted combustion.\n\nBiography:\n\nTaaresh
Taneja is a 5th year PhD candidate at the University of Minnesota (UMN)\,
Twin Cities\, who is currently focused on modeling non-equilibrium plasma
assisted combustion. He works with Prof. Suo Yang\, in the Computational R
eactive Flow and Energy Laboratory (CRFEL) at UMN. For his research\, he h
as received the UMII MNDrive Fellowship\, NSF INTERN Supplemental Funding
Opportunity and the UMN Doctoral Dissertation Fellowship\, along with othe
r travel grants for presenting his work at conferences. Taaresh has also i
nterned at the National Renewable Energy Laboratory\, Colorado and at the
Sandia National Laboratory\, California during his PhD. Before joining UMN
for his PhD in 2019\, Taaresh worked as a CFD engineer at Johnson Control
s India Pvt. Ltd.\, Pune (2017 - 2019). He received his B.E. (Hons) in Mec
hanical Engineering from BITS Pilani\, Goa Campus in 2017.\n\n\n\nHost Fac
ulty: Dr. Konduri Aditya\n\n\n\nAll are welcome
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:45@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240311T090000
DTEND;TZID=Asia/Kolkata:20240311T100000
DTSTAMP:20240306T051313Z
URL:https://cds.iisc.ac.in/events/seminar-cds-online-11th-march-computatio
nal-profiling-of-microbes-in-environments-pitfalls-controversies-and-solut
ions/
SUMMARY:{Seminar} @ CDS: Online : 11th March: "Computational profiling of m
icrobes in environments: pitfalls\, controversies\, and solutions"
DESCRIPTION:Department of Computational and Data Sciences\n\nDepartment Sem
inar\n\n\n\nSPEAKER : Dr. Jim Shaw\, Department of Mathematics at the
University of Toronto\nTITLE : "Computational profiling
of microbes in environments: pitfalls\, controversies\, and solutions"\nDa
te &\; Time : March 11\, 2024\, 09:00 AM\n\nVenue
: The Seminar will be held on MICROSOFT TEAMS. Please click on the fol
lowing link to join the Seminar:\n\nhttps://teams.microsoft.com/l/meetup-j
oin/19%3ameeting_NGZkMjVhNGYtNzZkMS00OGE3LTljNzItOTdhZTM1ZDMxNTdm%40thread
.v2/0?context=%7b%22Tid%22%3a%226f15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c
%22Oid%22%3a%220df9a69b-76ed-46bd-a19d-e01995f351a6%22%7d\n\n\n\nABSTRACT\
n\nDNA sequencing technology now allows researchers to detect the microorg
anisms in your microbiomes\, such as your gut\, skin\, and even blood. Com
putational methods for analyzing this data have enabled insights into the
relationship between our microbiomes and human health. DNA sequencing data
\, however\, is notoriously messy\; computational methods are also imperfe
ct.\n\nIn this talk\, I will first discuss how extremely subtle algorithmi
c issues have led to controversy in the biological literature\, including
suspicion about a landmark paper in the study of cancer microbiomes. After
giving an overview of the computational issues surrounding this field\, I
will present our new computational method for profiling microbiomes calle
d sylph (Shaw and Yu\, 2023\, bioRxiv). While sylph has high accuracy and
is >\; 30 times faster than previous methods\, I will focus on why our
new method works. More generally\, I will discuss the process of designing
practical algorithms for complicated biological data\, which I believe is
a treacherous but extremely rewarding process.\n\nBIOGRAPHY\n\nJim Shaw i
s a PhD candidate in the Department of Mathematics at the University of To
ronto under the supervision of Prof. Yun William Yu. His research centers
on developing efficient software and methods for computational biology fro
m a mathematical point of view\, with a heavy emphasis on bridging the gap
between theory and biological discovery. Jim is supported by an NSERC CGS
-D\, and prior to his PhD\, he received a BASc in Engineering Physics and
Mathematics from the University of British Columbia.\n\nHost Faculty: Dr.
Chirag Jain\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:42@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240306T103000
DTEND;TZID=Asia/Kolkata:20240306T113000
DTSTAMP:20240227T061300Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-defense-cds-06-ma
rch-2024-a-co-kurtosis-tensor-based-featurization-of-chemistry-for-scalabl
e-combustion-simulations/
SUMMARY:M.Tech Research: Thesis Defense: CDS: 06\, March 2024 "A co-kurtosi
s tensor based featurization of chemistry for scalable combustion simulati
ons"
DESCRIPTION:\n\nDEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\nM.Tech Res
earch Thesis Defense\n\n\n\nSpeaker : Mr. Dibya Jyoti Naya
k\nS.R. Number : 06-18-01-10-22-21-1-19747.\nTitle
: "A co-kurtosis tensor based featurization of chemistry for scalable c
ombustion simulations."\nResearch Supervisor : Dr. Konduri Aditya\nDate &a
mp\; Time : March 06\, 2024 (Wednesday)\, 10:30 am\nVenue
: Room No. 102 (CDS Seminar Hall)\n\n\n\n\nABSTRACT\nIdentifying l
ow-dimensional representations of the thermo-chemical state space for turb
ulent reacting flow systems is vitally important\, primarily to significan
tly reduce the computational cost of device-scale combustion simulations.
Moreover\, these simulations are often performed to gain fundamental insig
hts into the inception of extreme/anomalous events such as flashbacks\, fl
ame extinction\, blow-offs\, thermoacoustic instabilities\, etc.\, which c
an have detrimental effects on combustion efficiency and engine performanc
e. With the scale of scientific investigations ever increasing\, the need
for robust anomaly detection methods becomes increasingly critical for jud
icious steering of these simulations and also aiding smooth operations of
practical engines. Recent studies have shown that the fourth-order joint s
tatistical moment tensor\, i.e.\, co-kurtosis\, effectively captures anoma
lies/outliers in scientific data. Accordingly\, the primary objective of t
his work centers around leveraging the unique properties of the co-kurtosi
s tensor to drive low-cost and scalable combustion simulations and build r
obust algorithms for extreme event detection. Particularly\, the first par
t of this work develops tools for dimensionality reduction for chemistry\,
while the second part focuses on employing a co-kurtosis based detection
algorithm for capturing extreme events such as flame instabilities in hydr
ogen-fired reheat burners relevant to sequential gas turbine engines.\n\nT
o obtain low-dimensional manifolds (LDMs) that describe the original therm
o-chemical state\, principal component analysis (PCA) and its variants are
widely employed. An alternative dimensionality reduction technique that f
ocuses on higher order statistics\, co-kurtosis PCA (CoK-PCA)\, has been s
hown to provide an optimal LDM for effectively capturing the stiff chemica
l dynamics associated with spatiotemporally localized reaction zones. Whil
e its effectiveness has only been demonstrated based on a priori analyses
with linear reconstruction\, in this work\, we employ nonlinear techniques
to reconstruct the full thermo-chemical state and evaluate the efficacy o
f CoK-PCA compared to PCA. Specifically\, we combine a CoK-PCA-/PCA-based
dimensionality reduction (encoding) with an artificial neural network (ANN
) based reconstruction (decoding) and examine\, a priori\, the reconstruct
ion errors of the thermo-chemical state. We employ three combustion test c
ases representing varying degrees of complexity in the geometrical domain\
, combustion regimes\, ignition kinetics\, etc.\, to assess CoK-PCA/PCA co
upled with ANN-based reconstruction. Results from the analyses demonstrate
the robustness of the CoK-PCA based LDM with ANN reconstruction in accura
tely capturing the data\, specifically from the reaction zones.\n\nHydroge
n's highly reactive and diffusive nature towards decarbonization is prone
to flashbacks\, flame instabilities\, and thermoacoustic instabilities. Fo
r example\, in the case of reheat burners of hydrogen-fired sequential gas
turbine engines\, intermittent temperature and pressure fluctuations resu
lt in flame instabilities\, such as intermittent autoignition events at of
f-design locations that can adversely impact the engine's performance. To
address this issue\, we develop an unsupervised learning methodology based
on the co-kurtosis tensor to detect the early onset of spontaneous igniti
on kernels in lean premixed hydrogen combustion at vitiated conditions. Th
e accuracy of the model is evaluated for various ignition test cases.\n\n\
n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:40@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240301T150000
DTEND;TZID=Asia/Kolkata:20240301T160000
DTSTAMP:20240221T085040Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-colloquium-cds-sc
alable-read-alignment-algorithm-for-cyclic-pangenome-graphsscalable-read-a
lignment-algorithm-for-cyclic-pangenome-graphs/
SUMMARY:M.Tech Research Thesis {Colloquium}: CDS : "Scalable Video Data Man
agement and Visual Querying for Autonomous Camera Networks"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research
Thesis Colloquium\n\n\n\nSpeaker : Ms. Bharati Khanijo\n\nS.R. Number : 06
-18-02-10-12-19-1-17219\n\nTitle :"Scalable Video Data Management and Visu
al Querying for Autonomous Camera Networks"\nResearch Supervisor: Prof. Yo
gesh Simmhan\nDate &\; Time : March 01\, 2024 (Friday) at 03:00 PM\nVen
ue : # 102 CDS Seminar Hall\n\n\n\nABSTRACT\n\nVideo data has been histori
cally known for its unstructured nature and rich semantic content but also
for scalability issues in terms of storage and analytics. Mobile aerial p
latforms like drones capture such videos across space and time. Advances i
n computer vision and deep learning enable automatic extraction of rich se
mantic information from video data\, leading to applications where the sto
red video data can be used to study and analyze the world retrospectively
and automatically. However\, recent research has highlighted the compute-i
ntensive nature of such Deep Neural Network (DNN) models\, e.g.\, for accu
rate object detection\, leading to high computing costs that limits their
applicability for brute-force analysis of all historical videos. Also\, an
efficient design of such applications often requires co-analysis of video
data along with associated geospatial and temporal metadata\, which is a
challenge.\n\nWe propose a geospatial-temporal video query system with sup
port for semantic queries for drone videos\, extending an existing spatial
-temporal database and contemporary object detection models. We develop a
heuristic to enable better reuse of semantic object detections obtained fr
om different configurations (object detection model and its input resoluti
on) . The system further motivates the need for optimizations for retrospe
ctive semantic analysis and storage for drone videos\, which is addressed
by our novel DDownscale method and the associated ingest pipeline.\n\nPrio
r optimizations on semantic querying over video data focus on static camer
as from city-scale traffic/surveillance camera networks\, often leveraging
the spatial and temporal characteristics of associated videos\, which are
absent in videos recorded by mobile drone cameras. We specifically focus
on two such characteristics of drone videos. One is that drone videos have
shorter durations\, unlike those captured by static cameras. Another is t
hat there can be large variations in the level of detail of information ca
ptured across a fleet of drone cameras due to differences in the resolutio
n of the camera\, the altitude\, and the orientation from which the videos
were captured.\n\nSpecifically\, we address the need to intelligently sca
le-down the spatial resolution of videos to reduce the video storage costs
and semantic query/inferencing time. However\, conventional methods of ma
nual or profiling-based estimation of the ideal scaling ratio are compute-
intensive and/or time consuming for such heterogeneous feeds. We propose D
Downscale\, a novel method to dynamically select the downscale factor for
a video by utilizing the information on the object size in the video. We m
odel the downscale factor and associated drop in relative recall due to do
wnscaling as a function of object size in the downscaled video and demonst
rated that for a given DNN model and class of interest\, DDownscale genera
lizes well to the evaluated datasets. A DDownscale inequality between the
relative recall drop and the hyper-parameters of the method is derived. Th
is satisfies 98% of the dynamically downscaled videos across datasets\, ob
jects of interest and parameters. The algorithm achieve over 19% reduction
in total object detection time and 24% reduction in storage on average co
mpared to the baseline of storing/inferencing at the original resolution \
, for different user-specified target reduction in recall values ranging f
rom 1--30%\, and 96% of the downscaled videos are within the target recall
drop.\n\nA simpler specification at the time of ingest of target level of
detail (average ground spatial distance) captured in the video and the ha
rmonic mean of relative recall drop for the class of smallest object of in
terest and selected object detection model was derived using the above mod
eling to aid in the selection of a target level of detail. Additionally\,
we develop an ingest pipeline that reduces the time to ingest drone videos
using this dynamically downscaling method over heterogeneous edge acceler
ators\, and reduce the average turnaround time to ingest data from multipl
e clients by ~ 66%\, despite the downscaling time overhead\, compared to u
ploading original resolution video without downscaling.\n\n\n\nALL ARE WEL
COME
CATEGORIES:Events,MTech Research Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:39@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240227T140000
DTEND;TZID=Asia/Kolkata:20240227T150000
DTSTAMP:20240216T092452Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-27th-february-towards-se
cure-interpretable-and-scalable-machine-learning-applications-in-cyber-phy
sical-systems/
SUMMARY:{Seminar} @ CDS: #102 :27th February: "Towards Secure\, Interpretab
le\, and Scalable Machine Learning Applications in Cyber-Physical Systems"
DESCRIPTION:Department of Computational and Data Sciences\n\nDepartment Sem
inar\n\n\n\nSpeaker : Dr. Shailja Thakur\, Postdoctoral fellow\, New York\
n\nTitle : "Towards Secure\, Interpretable\, and Scalable Machine Learning
Applications in Cyber-Physical Systems"\n\nDate &\; Time : February 27
\, 2024\, 02:00 PM\n\nVenue : # 102\, CDS Seminar Hall\n\n\n\nABSTRACT\n\n
In a world increasingly reliant on Cyber-Physical Systems (CPS)\, there ar
e critical challenges associated with the integration of complex software
and hardware. The enormous and diverse nature of data\, alongside pressing
security and privacy concerns\, demands innovative solutions. My work aim
s to enhance the intelligence of CPS through AI\, aiming for systems that
are not only self-aware but also capable of adapting in real-time to chang
ing environments. To that end\, my work has spanned the automotive\, energ
y\, and hardware sectors\, delivering practical solutions engineered along
side industry partners. I have made significant strides in enhancing secur
ity in automotive systems and have pioneered tools for deciphering the dec
ision-making processes of machine learning models. In the realm of hardwar
e design\, I am exploring the potentials of Large Language Models (LLMs) t
o automate and optimize the process\, reducing human error and increasing
efficiency. In the future\, I want to expand upon the challenges and scope
of applying generative AI in CPS for developing time-efficient\, scalable
\, safe and transparent real-world applications.\n\nBIOGRAPHY\n\nI am a po
stdoctoral fellow currently at New York University in the Tandon School of
Engineering within the Department of Electrical and Computer Engineering
and the Center of Cyber Security with Professors Ramesh Karri and Professo
r Siddharth Garg. My research interests span the field of cybersecurity\,
with a particular focus on the application of language modelling in embedd
ed systems\, LLM attributions\, safety\, and privacy. Shailja received her
Ph.D from the University of Waterloo and a masters in Computer Science an
d Engineering from IIIT Delhi.\n\nHost Faculty: Dr. Danish Pruthi\n\n\n\nA
LL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:41@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240226T113000
DTEND;TZID=Asia/Kolkata:20240226T123000
DTSTAMP:20240226T055854Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-26th-february-domain-ada
ptation-for-fair-and-robust-computer-vision/
SUMMARY:{Seminar} @ CDS: #102 :26th February: "Domain Adaptation for Fair a
nd Robust Computer Vision."
DESCRIPTION:Department of Computational and Data Sciences\n\nCDS / KIAC Sem
inar\n\n\n\nSpeaker : Mr. Tarun Kalluri\, University of California San Die
go.\n\nTitle : "Domain Adaptation for Fair and Robust Computer Vision."\n\
nDate &\; Time : February 26\, 2024\, 11:30 AM\n\nVenue : # 102\, CDS S
eminar Hall\n\n\n\nABSTRACT\n\nWhile recent progress significantly advance
s the state of the art in computer vision across several tasks\, the poor
ability of these models to generalize to domains and categories under-repr
esented in the training set remains a problem\, posing a direct challenge
to fair and inclusive computer vision. In my talk\, I will talk about my r
ecent efforts towards improving generalizability and robustness in compute
r vision using domain adaptation. First\, I will talk about our work on sc
aling domain adaptation to large scale datasets using metric learning. Nex
t\, I will introduce our new dataset effort called GeoNet aimed at benchma
rking and developing novel algorithms towards geographical robustness in v
arious vision tasks. Finally\, I will talk about the latest research study
ing the role of language supervision to improve adaptation of visual model
s to new domains.\n\nBIOGRAPHY\n\nTarun Kalluri is a fifth year PhD studen
t at UC San Diego in the Visual Computing Group. Prior to that\, he gradua
ted with a bachelors from Indian Institute of Technology\, Guwahati and wo
rked as a data scientist in Oracle. His research interests lie in label an
d data efficient learning from images and videos\, domain adaptation and i
mproving fairness in AI. He is a recipient of IPE PhD fellowship.\n\nHost
Faculty: Prof. Venkatesh Babu\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:36@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240215T143000
DTEND;TZID=Asia/Kolkata:20240215T154500
DTSTAMP:20240209T155828Z
URL:https://cds.iisc.ac.in/events/cds-seminar-systemsai-research-microsoft
/
SUMMARY:[CDS Seminar] Systems+AI Research @ Microsoft
DESCRIPTION:Department of Computational and Data Sciences (CDS) Seminar\n\n
\n\nTitle: Systems + AI Research @ Microsoft\n\nSpeakers: Saravan Rajmohan
\, Chetan Bansal\, Anjaly Parayil and Mayukh Das\, Microsoft\n\nDate/Time:
Thu Feb 15\, 230-345PM\n\nVenue: SERC Auditorium 4th Floor\n\n\n\nAbstrac
t: At Microsoft\, we operate one of the largest productivity clouds and we
need to keep pace with paradigm shifts such as the massive growth in AI w
orkloads\, sustainability push\, the need for self-managing cloud environm
ents and the complex challenges that arise out of its sheer scale. To solv
e these challenges\, at M365 Research group in Microsoft\, we have built a
cross-domain research team focusing on applied research on “ML for Syst
ems” to bring a step function improvement in Cloud Efficiency and Reliab
ility. The group comprises of 30+ PhD Researchers\, Research Fellows and I
nterns in Bangalore (co-located with MSR India)\, Redmond and Beijing. We
aim to partner with top research institutions to drive innovation and leve
rage the immense scientific knowledge and expertise to bring new ideas int
o practice. Our goal is to explore all the dimensions from collaborative r
esearch with faculty members\, to establishing knowledge sharing seminars
and building a long-term talent pipeline. You can read more about our rese
arch group here: https://aka.ms/systems-innovation.\n\nBios:\n\n Saravan
Rajmohan is a Partner Director at Microsoft\, leading the M365 Research g
roup focused on AI and Applied Research. He oversees diverse teams in the
US\, Latin America\, UK\, India\, and China\, conducting groundbreaking re
search in Systems Innovation and Privacy Preserving Machine Learning. The
group develops advanced algorithms and hardware innovation to enhance M365
infrastructure and services\, collaborating closely with Microsoft Resear
ch labs. They drive AIOps Cloud Intelligence research\, improving efficien
cy and reliability across M365 services. His Efficient AI team optimizes g
enerative AI scenarios through cross-layer optimization and collaboration\
, while privacy-preserving research ensures confidentiality in ML systems\
, playing a pivotal role in AI privacy policies.\n Chetan Bansal is a Pri
ncipal Research Manager at Microsoft leading the AIOps and Cloud Intellige
nce research. His research is focused on improving Cloud Efficiency and Re
liability through AI and data-driven methods. His research has had a signi
ficant impact on Cloud Efficiency and Developer Productivity at Microsoft
with millions of dollars of savings. He has published 30+ papers in top in
ternational conferences like ICSE\, FSE\, NSDI\, KDD and filed 15+ patents
. His research has been recognized with best paper awards and nominations
in conferences such as SoCC\, FMCAD and ICSE. Prior to Microsoft\, his wor
k on Formal verification of Web Protocols was recognized by awards from Fa
cebook\, Mozilla and Google.\n Anjaly Parayil is a Senior Researcher at M
365 Research leading applied research at the intersection of efficiency an
d reliability of cloud services. In particular\, she works on data drive
n optimizations to ensure continuous availability of cloud services as wel
l as workload aware techniques for improving the efficiency of Cloud infra
structure. Before joining Microsoft\, Anjaly was a Post Doc with the Compu
tational and Information Sciences Directorate at the US Army Research Labo
ratory where she focused on Reinforcement Learning and Bayesian Inferencin
g. Anjaly completed her graduate studies at the Department of Aerospace En
gineering at IISc with her thesis focusing on uncertain systems and multi-
agent control\, for which she received Prof. A. K. Rao Medal for the best
Ph.D. Thesis. Anjaly has authored 25+ publications in Artificial intellige
nce\, control systems and optimization.\n Mayukh Das is a Senior Research
er at Microsoft driving applied AI research for Cloud Efficiency. In par
ticular\, he works on varied decision-making problems for configuration tu
ning for performance optimization of cloud services\, for capacity provisi
oning\, for power and energy optimization\, and\, operational efficiency o
f ML workloads. He completed his PhD from UT Dallas and his thesis work wa
s focused on Guided Reinforcement Learning and Probabilistic Modeling in N
oisy domains. Prior to Microsoft he was at Samsung Research solving Edge-A
I problems. He serves on the program committee of various conferences incl
uding AAAI\, ICML\, NeurIPS\, SDM etc. and has served as a track chair at
CODS-COMAD ‘24. Mayukh has authored 25+ publications in AI/ML and holds
7+ patents.\n\n\n\n\nHost: Yogesh Simmhan\, CDS\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:37@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240213T140000
DTEND;TZID=Asia/Kolkata:20240213T150000
DTSTAMP:20240209T160037Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-13th-february-on-using-m
achine-learning-techniques-for-the-numerical-solution-of-convection-diffus
ion-problems/
SUMMARY:{Seminar} @ CDS: #102 :13th February: "On Using Machine Learning Te
chniques for the Numerical Solution of Convection-Diffusion Problems"
DESCRIPTION:Department of Computational and Data Sciences\n\nDepartment Sem
inar\n\n\n\nSpeaker : Prof. Dr. Volker John\, Berlin Germany\n\nTitle : "O
n Using Machine Learning Techniques for the Numerical Solution of Convecti
on-Diffusion Problems"\n\nDate &\; Time : February 13\, 2024\, 02:00 PM
\n\nVenue : # 102\, CDS Seminar Hall\n\n\n\nABSTRACT\n\nIn the convection-
dominated regime\, solutions of convection-diffusion problems usually poss
esses layers\, which are regions where the solution has a steep gradient.
It is well known that many classical numerical discretization techniques f
ace difficulties when approximating the solution to these problems.\n\nMan
y stabilized discretizations of steady-state convection-diffusion equation
s lead to numerical solutions with notable spurious oscillations in a vici
nity of layers. Using discontinuous Galerkin methods offers the possibilit
y to reduce such oscillations effectively with post-processing techniques\
, so-called slope limiters. In the first part of the talk\, several of the
se techniques from the literature and improvements proposed in [1\, 2] wil
l be discussed and numerical assessments will be presented. Then the quest
ion will be studied whether a feed forward neural network (multilayer perc
eptrons)\, which is trained on the basis of these techniques\, is likewise
able to limit spurious oscillations\, see [3].\n\nIn recent years\, physi
cs-informed neural networks (PINNs) for approximating the solution to (ini
tial-)boundary value problems received a lot of interest. In the second pa
rt of this talk\, various loss functionals for PINNs are studied that are
novel in the context of PINNs and are especially designed for convectiondo
minated convection-diffusion problems. They are numerically compared to th
e vanilla and a hp-variational loss functional from the literature based o
n two benchmark problems whose solutions possess different types of layers
. It is observed in [4] that the best novel loss functionals reduce the L
2 (Ω) error by 17.3% for the first and 5.5% for the second problem compa
red to the methods from the literature.\n\nThis talk presents joint work w
ith Derk Frerichs–Mihov and Linus Henning\n\nReferences\n\n[1] D. Freric
hs\, V. John\, On reducing spurious oscillations in discontinuous Galerkin
(DG) methods for steady-state convection–diffusion equations\, Journal
of Computational and Applied Mathematics 393 (2021) 113487. doi:10.1016/j.
cam.2021.113487.\n\n[2] D. Frerichs-Mihov\, V. John\, On a technique for r
educing spurious oscillations in DG solutions of convection–diffusion eq
uations\, Applied Mathematics Letters 129 (2022) 107969. doi:10.1016/j.aml
.2022.107969.\n\n[3] D. Frerichs-Mihov\, L. Henning\, V. John\, Using deep
neural networks for detecting spurious oscillations in discontinuous Gale
rkin solutions of convection-dominated convection-diffusion equations\, J.
Sci. Comp. 97 (2023)\, Article 36. doi.org/10.1007/s10915-023-02335-x.\n\
n[4] D. Frerichs-Mihov\, L. Henning\, V. John\, On loss functionals for ph
ysicsinformed neural networks for convection-dominated convection-diffusio
n problems\, WIAS Preprint 3063\, submitted\, 2023 doi.org/10.20347/ WIAS.
PREPRINT.3063.\n\nBIOGRAPHY\nUniv.-Prof. Dr. Volker John - Head of Researc
h Group Numerical Mathematics and Scientific Computing Weierstrass Institu
te for Applied Analysis and Stochastics Mohrenstr. 39 10117 Berlin\, Germa
ny\nMore details about Prof. Volker John is available at https://www.wias-
berlin.de/people/john/\n\nHost Faculty: Prof. Sashikumaar Ganesan\n\n\n\nA
LL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:38@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240213T113000
DTEND;TZID=Asia/Kolkata:20240213T123000
DTSTAMP:20240209T160253Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-13th-february-large-scal
e-real-space-finite-difference-based-electronic-structure-calculations/
SUMMARY:{Seminar} @ CDS: #102 :13th February: "Large-scale real-space finit
e-difference based electronic structure calculations"
DESCRIPTION:Department of Computational and Data Sciences\n\nDepartment Sem
inar\n\n\n\nSpeaker : Dr. Abhiraj Sharma\, Postdoctoral Researcher\, LBNL\
, California\, U.S.A.\n\nTitle : "Large-scale real-space finite-difference
based electronic structure calculations"\n\nDate &\; Time : February 1
3\, 2024\, 11:30 AM\n\nVenue : # 102\, CDS Seminar Hall\n\n\n\nABSTRACT\n\
nElectronic structure methods are widely employed for understanding the qu
antum mechanics driven behavior of material systems including but not limi
ted to flexo electricity\, spintronics\, piezoelectricity\, and thermal co
nductivity. Among these methods\, Kohn-Sham density functional theory (DFT
) has emerged as the cornerstone of material's research owing to its high
accuracy-to-cost ratio in comparison to other such ab initio methods. Desp
ite significant advancements\, Kohn-Sham calculations are still very expen
sive which limits the length and time scales accessible to first-principle
s investigation.\n\nIn this talk\, I will present on the recent developmen
ts in real-space finite-difference DFT resulting in a significant reductio
n in the computational cost and wall times associated with Kohn-Sham calcu
lations\, opening avenues for studying material systems considered intract
able before. In particular\, we have achieved more than an order of magnit
ude speedup in comparison to state-of-the-art DFT implementations by utili
zing the computational locality in finite-difference method\, concepts of
symmetry\, and high performance hybrid computing architectures. I will als
o discuss our recent efforts on accelerating the quantum molecular dynamic
s (QMD) simulations by incorporating a kernel method based machine-learned
force field (MLFF) scheme trained in an online manner\, resulting in more
than two orders of magnitude speedups. Finally\, I will discuss the compu
tational bottlenecks associated with calculating higher order properties o
f material systems including superconductivity and thermal conductivity an
d our ongoing work on developing a large-scale real-space formulation of d
ensity functional perturbation theory (DFPT) to overcome it\, opening aven
ues for a plethora of new applications with technological implications.\n\
nBIOGRAPHY\n\nAbhiraj Sharma received his B.Tech degree in Civil Engineeri
ng from IIT Roorkee in 2016. Afterwards\, he completed his M.S. and PhD fr
om Georgia Institute of Technology in 2022. Currently\, he is working as a
Postdoctoral Researcher in the Physical and Life Sciences division at Law
rence Livermore National Lab. His research is in the broad area of materia
l physics and mechanics with the focus on the development of mathematical
and computational tools to enable the first-principles study of mechanics
in material which can potentially lead to the discovery of materials with
fascinating properties.\n\nHost Faculty: Dr. Phani Motamarri\n\n\n\nALL AR
E WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:35@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240207T100000
DTEND;TZID=Asia/Kolkata:20240207T110000
DTSTAMP:20240207T172508Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-cds-a-scalable-as
ynchronous-discontinuous-galerkin-method-for-massively-parallel-pde-solver
s/
SUMMARY:Ph.D. Thesis {Colloquium}: CDS : "A scalable asynchronous discontin
uous-Galerkin method for massively parallel PDE solvers."
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Col
loquium\n\n\n\n\n\n\n\n\n\n\n\nSpeaker : Mr. Shubham Kumar Goswami\n\nS.
R. Number : 06-18-00-10-12-19 -1-17224\n\nTitle : "A scalable asynchro
nous discontinuous-Galerkin method for massively parallel PDE solvers "\n\
nResearch Supervisor: Dr. Konduri Aditya\nDate &\; Time : February
07\, 2024 (Wednesday) at 10:00 AM\nVenue : The Thesis Colloquium will b
e held on HYBRID Mode\n# 102 CDS Seminar Hall /MICROSOFT TEAMS.\n\n\n\n\nP
lease click on the following link to join the Thesis Colloquium:\n\nMS T
eams link\n\n\n\n\n\n\nABSTRACT\n\n\n\n\n\n\n\n\n\nAccurate simulations of
turbulent flows in computational fluid dynamics (CFD) are crucial for com
prehending numerous complex phenomena in engineered systems and natural pr
ocesses. These flows are governed by nonlinear partial differential equati
ons (PDEs)\, which are approximated as algebraic equations and solved usin
g PDE solvers. However\, the complexity of turbulence makes these simulati
ons computationally expensive\, necessitating the use of massively paralle
l supercomputers. While advancements such as hardware-aware computing\, fa
ult tolerance\, and overlapping computation and communication have improve
d solver scalability\, achieving efficient performance at extreme scales r
emains a challenge owing to the communication and synchronization overhead
. To address this issue\, an asynchronous computing approach was introduce
d that relaxed communication and synchronization at a mathematical level\,
allowing PEs to operate independently regardless of the status of message
s\, potentially decreasing communication overhead and enhancing scalabilit
y. This approach has been developed specifically for finite difference sch
emes\, which are widely used but not ideal for complex geometries and unst
ructured meshes. The objective of this work is to develop an asynchronous
discontinuous-Galerkin method that can provide high-order accurate solutio
ns for various flow problems on unstructured meshes and demonstrate its sc
alability.\n\n\n\n\nBased on the asynchronous computing approach\, several
PDE solvers have been developed that use high-order asynchrony-tolerant f
inite difference schemes for spatial discretization to simulate reacting a
nd non-reacting turbulent flows\, achieving significant improvements in sc
alability. However\, for time integration\, most of them used either multi
-step Adams-Bashforth schemes\, which possess poor stability\, or multi-st
age Runge-Kutta (RK) schemes with an over-decomposed domain that necessita
tes larger message sizes for communication and redundant computations. In
this work\, we propose a novel method to couple asynchrony-tolerant and lo
w-storage explicit RK (LSERK) schemes to solve time-dependent PDEs with re
duced communication efforts. We developed new asynchrony-tolerant schemes
for ghost or buffer point updates that are necessary to maintain the desir
ed order of accuracy. The accuracy of this method has been investigated bo
th theoretically and numerically using simple one-dimensional linear model
equations. Thereafter\, we demonstrate the scalability of the proposed nu
merical method through three-dimensional simulations of decaying Burgers
’ turbulence performed using two different asynchronous algorithms: comm
unication-avoiding and synchronization-avoiding algorithms. Scalability st
udies up to 27\,000 cores yielded a speed-up of up to 6× compared to a ba
seline synchronous algorithm.\n\n\n\n\nIn recent years\, the discontinuous
Galerkin (DG) method has received broad interest in developing PDE solver
s\, particularly for nonlinear hyperbolic problems\, due to its ability to
provide high-order accurate solutions in complex geometries\, capture dis
continuities\, and exhibit high arithmetic intensity. However\, the scalab
ility of DG-based solvers is hindered by communication bottlenecks that ar
ise at extreme scales. In this work\, we introduce the asynchronous DG (AD
G) method\, which combines the benefits of the DG method with asynchronous
computing by relaxing the need for data communication and synchronization
at the mathematical level to overcome communication bottlenecks. The prop
osed ADG method ensures flux conservation and effectively addresses challe
nges arising from asynchrony. To assess its stability\, we employ Fourier-
mode analysis to examine the dissipation and dispersion behavior of fully-
discrete DG and ADG schemes with the Runge-Kutta (RK) time integration sch
emes across the entire range of wavenumbers. Furthermore\, we present an e
rror analysis within a statistical framework\, which demonstrates that the
ADG method with standard numerical fluxes achieves at most first-order ac
curacy. To recover accuracy\, we derived asynchrony-tolerant (AT) fluxes t
hat utilize data from multiple time levels. Finally\, extensive numerical
experiments are conducted to validate the performance and accuracy of the
ADG-AT scheme for both linear and nonlinear problems.\n\n\n\n\nWith the de
velopment of the asynchronous discontinuous-Galerkin (ADG) method\, we fin
ally put our focus on implementing and evaluating its performance in solvi
ng hyperbolic equations with shocks/discontinuities. To achieve this\, we
chose a highly scalable DG solver for compressible Euler equations from de
al.II\, which is one of the widely used open-source finite element librari
es. The solver uses low-storage explicit Runge-Kutta schemes for the time
integration. We implemented the ADG method in deal.II\, incorporating the
communication-avoiding algorithm (CAA)\, and performed validation and benc
hmarking\, showcasing the accuracy limitations of standard ADG schemes and
the effectiveness of newly developed asynchrony-tolerant (AT) fluxes. Str
ong scaling results are provided for both synchronous and asynchronous DG
solvers\, demonstrating a speedup of up to 80%. Since these AT fluxes are
also compatible with the finite volume (FV) method\, the overall work high
lights the potential benefits of the asynchronous approach for the develop
ment of accurate and scalable DG and FV-based PDE solvers\, paving the way
for simulations of complex physical systems on massively parallel superco
mputers.\n\n\n\n\n\n\n\nALL ARE WELCOME\n\n
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:34@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240131T140000
DTEND;TZID=Asia/Kolkata:20240131T150000
DTSTAMP:20240126T064354Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-colloquium-cds-sc
alable-read-alignment-algorithm-for-cyclic-pangenome-graphs/
SUMMARY:M.Tech Research Thesis {Colloquium}: CDS : "Scalable read alignment
algorithm for cyclic pangenome graphs"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research
Thesis Colloquium\n\n\n\nSpeaker : Ms. Jyotshna Rajput\n\nS.R. Number : 06
-18-01-10-22-21-1-19943\n\nTitle :"Scalable read alignment algorithm for c
yclic pangenome graphs"\nResearch Supervisor: Dr. Chirag Jain\nDate &\;
Time : January 31\, 2024 (Wednesday) at 02:00 PM\nVenue : # 102 CDS Semin
ar Hall\n\n\n\nABSTRACT\n\nThe growing availability of genome sequences of
several species\, including humans\, has created the opportunity to utili
ze multiple reference genomes for bioinformatics analyses and improve the
accuracy of genome resequencing workflows. Graph-based data structures are
suitable for compactly representing multiple closely related reference ge
nomes. Pangenome graphs use a directed graph format\, where vertices are l
abeled with strings\, and the individual reference genomes are represented
as paths in the graph. Aligning sequences (reads) to pangenome graphs is
fundamental for pangenome-based genome resequencing. The sequence-to-graph
alignment problem seeks a walk in the graph that spells a sequence with m
inimum edit distance from the input sequence. However\, exact algorithms f
or solving this problem are unlikely to scale due to the known hardness of
this problem. Co-linear chaining is a well-studied and commonly used heur
istic alternative for quickly aligning reads to a graph. However\, the kno
wn chaining algorithms are restricted to directed acyclic graphs (DAGs) an
d are not trivially generalizable to cyclic graphs. This limitation must b
e addressed because pangenome graphs often contain cycles due to inversion
s\, duplications\, or copy number mutations within the reference genomes.\
n\nThis thesis presents the first practical formulation and algorithm for
co-linear chaining on cyclic pangenome graphs. Our algorithm builds upon t
he known chaining algorithms for DAGs. We propose a novel iterative algori
thm to handle cycles and provide rigorous proof of correctness and runtime
complexity. Our algorithm also exploits the domain-specific small-width p
roperty of pangenome graphs. The proposed optimizations enable our algorit
hm to scale to large human pangenome graphs. We implemented our algorithm
in C++ and referred to it as PanAligner (https://github.com/at-cg/PanAlign
er). PanAligner is an end-to-end long-read aligner for pangenome graphs. W
e evaluated its speed and accuracy by aligning simulated long reads to a c
yclic human pangenome graph comprising 95 haplotypes. We achieved superior
read mapping accuracy using PanAligner compared to existing methods.\n\n\
n\nALL ARE WELCOME
CATEGORIES:Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:33@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240131T140000
DTEND;TZID=Asia/Kolkata:20240131T150000
DTSTAMP:20240126T064012Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-online-mode-cds-31-j
anuary-2024-sparsification-of-reaction-diffusion-dynamical-systems-on-comp
lex-networks/
SUMMARY:Ph.D: Thesis Defense: Online Mode: CDS: 31\, January 2024 "Sparsifi
cation of Reaction-Diffusion Dynamical Systems on > Complex Networks"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\nPh.D. Thesis D
efense\n\n\n\nSpeaker : Mr.Abhishek Ajayakumar\n\nS.R. Number : 06-18-01-1
0-12-18-1-16176\n\nTitle : "Sparsification of Reaction-Diffusion Dynamical
Systems on Complex Networks"\n\nResearch Supervisor : Prof. Soumyendu Rah
a\n\nDate &\; Time : January 31\, 2024 (Wednesday)\, 02:00 PM\nVenue :
The Thesis Défense will be held on MICROSOFT TEAMS\n\nPlease click on the
following link to join the Thesis Defense:\n\nMS Teams link\n\n\n\nABSTRA
CT\n\nGraph sparsification is an area of interest in computer science and
applied mathematics. Sparsification of a graph\, in general\, aims to redu
ce the number of edges in the network while preserving specific properties
of the graph\, like cuts and subgraph counts. Modern deep learning framew
orks\, which utilize recurrent neural network decoders and convolutional n
eural networks\, are characterized by a significant number of parameters.
Pruning redundant edges in such networks and rescaling the weights can be
useful. Computing the sparsest cuts of a graph is known to be NP-hard\, an
d sparsification routines exist for generating linear-sized sparsifiers in
almost quadratic running time.The complexity of this task varies\, closel
y linked to the desired level of sparsity to achieve. The thesis introduce
s the concept of sparsification to the realm of reaction-diffusion complex
systems. The aim is to address the challenge of reducing the number of ed
ges in the network while preserving the underlying flow dynamics. Sparsifi
cation of such complex networks is approached as an inverse problem guided
by data representing flows in the network\, where a relaxed approach is a
dopted considering only a subset of trajectories. The network sparsificati
on problem is mapped to a data assimilation problem on a reduced order mod
el (ROM) space with constraints targeted at preserving the eigenmodes of t
he Laplacian matrix under perturbations. The Laplacian matrix is the diffe
rence between the diagonal matrix of degrees and the graph’s adjacency m
atrix. Approximations are propose to the eigenvalues and eigenvectors of t
he Laplacian matrix subject to perturbations for computational feasibility
\, and a custom function is included based on these approximations as a co
nstraint on the data assimilation framework. Extensive empirical testing c
overed a range of graphs\, while its application to multiple instances led
to the creation of sparse graphs. In the latter phase of the thesis\, a f
ramework is presented to enhance proper orthogonal decomposition (POD)-bas
ed model reduction techniques in reaction-diffusion complex systems. This
framework incorporates techniques from stochasticfiltering theory and patt
ern recognition (PR). Obtaining optimal state estimates from a noisy model
and noisy measurements forms the core of the filtering problem. By integr
ating the particle filtering technique\, the reaction-diffusion state vect
ors are generated at various time steps\, utilizing the ROM states as meas
urements. To ensure the framework’s effectiveness\, intermittent updates
to the system variables are made during the particle filtering step\,empl
oying the carefully crafted sparse graph. The framework is utilized for ex
perimentation\, and results are presented on random graphs\, considering t
he diffusion equation on the graph and the chemical Brusselator model as t
he reaction-diffusion system embedded in the graph. Limitations of the met
hod are discussed\, and the proposed framework is evaluated by comparing i
ts performance against the Neural Ordinary Differential Equation or neural
ODE-based approach\, which serves as a compelling reference due to its de
monstrated robustness in specific applications.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:32@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240129T093000
DTEND;TZID=Asia/Kolkata:20240129T103000
DTSTAMP:20240119T062759Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-hybridcds-29-january
-2024end-to-end-resiliency-analysis-framework-for-cloud-storage-services/
SUMMARY:Ph.D: Thesis Defense: HYBRID:CDS: 29\, January 2024 "End-to-end Res
iliency Analysis Framework for Cloud Storage Services."
DESCRIPTION:\nDEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\n\nPh.D. Thes
is Defense \n\n\n\nSpeaker : Ms. Archita Ghosh\nS.R. Numbe
r : 06-18-02-10-12-17-1-14492\nTitle : "End-to-e
nd Resiliency Analysis Framework for Cloud Storage Services."\nResearch Su
pervisor :Dr Lakshmi Jagarlamudi\nDate &\; Time : January 29\, 2024
(Monday)\, 09:30 AM\nVenue : The Thesis Défense w
ill be held on HYBRID Mode\n\n# 102 CDS Seminar Hall /MICROSOFT TEAMS\n
\nPlease click on the following link to join the Thesis Defense:\nMS Teams
link\n\n\n\n\n\nABSTRACT\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\
n\n\n\n\n\n\nCloud storage service brought the idea of a global scale stor
age system available on-demand and accessible from anywhere. Despite the b
enefits\, resiliency remains one of the key issues that hinder the wide ad
aptation of storage services. The data is hosted on cloud data centers con
taining hundreds of thousands of commodity-grade hardware with layers of c
omplex software. Failures due to system crashes\, natural disasters\, cybe
r-attacks\, etc.\, are common and frequent in such environments. To keep t
he service unaffected by such events\, resiliency is essential for cloud s
ystems. For storage services\, resiliency is far more critical because los
ing access to data or\, more importantly\, a complete data loss can have a
catastrophic impact on the client.\n\n\n\n\n\n\n\n\nThe existing works on
storage resiliency focus on maintaining sufficient user data redundancy i
n the system to maintain a reliable service. However\, providing a global-
scale storage solution requires various functional and management layers t
o ensure the service is accessible and all the stored items are durable. T
he first part of our work proves that resiliency at the stored data level
does not guarantee service level reliability. A generic cloud storage syst
em model is designed to analytically show that the reliability achieved at
the service level drastically differs from the reliability ensured by sto
red data redundancy. This motivates us to bring the entire system into pur
view to understand cloud storage resiliency.\n\nDue to the complexity and
variation of large-scale storage architectures\, assessing end-to-end stor
age resiliency is a challenging task. To achieve this\, the second part of
the work proposes a generic resiliency evaluation method for cloud storag
e services. The method identifies the essential functional layers for stor
age service and the components constituting the layers. It then performs a
n in-depth behavior analysis during all possible failures of each componen
t. The method is used to assess the resiliency of two diverse and real-wor
ld cloud storage services\, OpenStack Swift and CephFS. The analysis ident
ifies various resiliency weak points in the service architectures and depi
cts the effectiveness of different resiliency methods used at various laye
rs.\n\nThe third part of the work extends the resiliency evaluation method
to understand the correlation of resiliency with the service usage patter
n. A storage service can be used for different use cases resulting in the
variation of request interarrival time\, read and write ratio\, accessed d
ata and metadata\, etc. Hence\, the components involved in access sequence
s may differ\, and so can their failure impact. Using the improved resilie
ncy evaluation method and access patterns identified from real traces\, we
show that resiliency can be selective and dynamically adjusted based on w
orkloads without affecting service reliability.\n\nFinally\, the work defi
nes an end-to-end resiliency analysis framework for cloud storage services
that enables quantification\, comparison\, and optimization of cloud stor
age resiliency. The framework allows effective modeling of cloud storage r
esilience by combining the resiliency of each component participating in s
ervice reliability maintenance for specific workloads. The framework succe
ssfully models the resiliency of OpenStack Swift and CephFS as Stochastic
Petri Nets (SPNs). The models are used to quantify and compare the resilie
ncy of the above two service architectures and demonstrate how to optimize
resiliency while achieving expected service reliability.\n\n\n\nALL ARE W
ELCOME\n\n\n\n\n\n
CATEGORIES:Events,Thesis Defense
END:VEVENT
BEGIN:VEVENT
UID:23@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata;VALUE=DATE:20240124
DTEND;TZID=Asia/Kolkata;VALUE=DATE:20240126
DTSTAMP:20231222T135636Z
URL:https://cds.iisc.ac.in/events/workshop-on-artificial-intelligence-in-p
recision-medicine/
SUMMARY:Workshop on Artificial Intelligence in Precision Medicine
DESCRIPTION:Date: 24-25 January\, 2024\n\nLocation: #102-Seminar Hall\, Co
mputational &\; Data Sciences Department\, IISc\n\n\n\nThe workshop aim
s to foster a deeper understanding of ongoing research in both Industry an
d Academia by bringing together leaders from both. We hope to provide a co
mprehensive overview of recent advancements in precision medicine\, and al
so explore its wide-ranging applications.\n\nFor more information\, please
visit our website: website\n\nMandatory Poster Submission\n\nIn line with
our commitment to fostering interactive discussions and knowledge exchang
e\, we require all attendees to submit posters presenting their research c
ontributions relevant to the workshop themes. Poster submissions are manda
tory for participation in the workshop.\n\nYou can submit your poster and
register yourself here: form\n\nNote: All the participants are requested t
o kindly carry your own laptops for the hands-on session.\n\n \;\n\n\n
\n\n\n \;
CATEGORIES:Events
END:VEVENT
BEGIN:VEVENT
UID:31@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240119T141500
DTEND;TZID=Asia/Kolkata:20240119T151500
DTSTAMP:20240118T150345Z
URL:https://cds.iisc.ac.in/events/cds-seminar-intl-collaborative-opportuni
ties-with-cardiff-university-on-cybersecurity-and-related-areas/
SUMMARY:[CDS SEMINAR] Intl Collaborative Opportunities with Cardiff Univers
ity on Cybersecurity and Related Areas
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\nDEPARTMENT SEM
INAR\n\n\n\nSpeaker: Dr. A. Javed\, Cardiff University\, UK\n\nTitle: In
ternational Collaborative Opportunities with Cardiff University on Cyberse
curity and Related Areas\n\nDate and Time: Fri Jan 19\, 2:15pm\n\nLocation
: CDS 102 Seminar Hall\n\n\n\nABSTRACT: \n\nThis talk will provide a com
prehensive overview of recent advancements and cutting-edge research we do
in the cybersecurity group at Cardiff University. Our researchers have be
en at the forefront of addressing contemporary challenges in cybersecurity
\, ranging from understanding human and machine behaviour behind cyber-att
acks to developing innovative models for threat detection and prediction.
The discussion will encompass a spectrum of topics\, including exploring n
ovel methods in community organisation within online social networks\, the
correlation between emotions and malware propagation\, and the developmen
t of effective countermeasures against evolving cyber threats. The talk ai
ms to outline our work and identify areas where we can collaborate and sub
mit joint grant applications.\n\nBIO: \n\nDr. A. Javed is a Cardiff Unive
rsity lecturer focusing on machine learning to identify cybercriminal beha
viour. His research includes understanding human and machine behaviour pat
terns\, emphasising online social network (OSN) security. Dr. Javed collab
orates with other schools\, resulting in joint publications on predicting
drive-by download attacks and factors in malware propagation. He pioneers
predictive defence strategies in OSN security. He also develops models for
securing Internet of Things (IoT) devices on the edge and addressing adve
rsarial attacks. His research extends to quantifying organisational cyber
risks for continuous monitoring\, aiding in resource allocation. Dr. Javed
's work enhances his teaching and contributes to the fields of cyber secur
ity and risk management.\n\nThis talk is supported by the Global Wales-IIS
c Joint Research Partnership Fund\n\n\n\nHOST: Yogesh Simmhan\n\nALL ARE W
ELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:29@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240112T160000
DTEND;TZID=Asia/Kolkata:20240112T170000
DTSTAMP:20240108T064154Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-colloquium-cds-an
-importance-sampling-in-n-sphere-monte-carlo-nsmc-and-its-performance-anal
ysis-for-high-dimensional-integration/
SUMMARY:M.Tech Research Thesis {Colloquium}: CDS : "An importance sampling
in N-Sphere Monte Carlo (NSMC) and its performance analysis for high dimen
sional integration."
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research
Thesis Colloquium\n\n\n\nSpeaker : Mr. Jawlekar Abhijeet Rajendra\n\nS.R.
Number : 06-18-01-10-22-21-1-20128\n\nTitle :"An importance sampling in N-
Sphere Monte Carlo (NSMC) and its performance analysis for high dimensiona
l integration"\n\nResearch Supervisor: Prof. Murugesan Venkatapathi\n\nDat
e &\; Time : January 12\, 2024 (Friday) at 04:00 PM\n\nVenue : # 102 CD
S Seminar Hall\n\n\n\nABSTRACT\n\nStatistical methods for estimating integ
rals are indispensable when the number of dimensions (parameters) become g
reater than ~ 10\, where numerical methods are unviable in general. Well
-known statistical methods like Quasi-Monte Carlo converge quickly only fo
r problems with a small number of effective dimensions\, and Markov Chain
Monte Carlo (MCMC) methods incur a sharply increasing computing effort wit
h the number of dimensions 'n' that is bounded as O(n^5). This bound on
the dimensional scaling of computing effort in multiphase MCMC is limited
to domains with a given convex shape (determined by the limits of integrat
ion). Note that the non-convexity and roughness of the boundaries of the
domain are factors that adversely affect the convergence of such methods
based on a random walk\, as the n-volume concentrates near the boundary in
high dimensions with increasing 'n'.\n\nA different approach to high-D in
tegration using (1D) line integrals along random directions coupled with a
less-known volume transformation was suggested here at the Institute by A
run et. al. This method called as N-sphere Monte Carlo (NSMC) is agnostic
to the shape and roughness of the boundary for any given distribution of e
xtents of the domain from a reference point. While the dimensional scaling
of computing in NSMC integration can be bound as O(n^3) for any distribut
ion of relative extents (and not a particular convex shape)\, a similar bo
und does not exist for MCMC as any given extent distribution can represent
numerous geometries where it may not converge. It was shown earlier tha
t when restricted to convex shapes where the extent density functions beco
me increasingly heavy tailed as 'n' increases\, the naive NSMC may be more
efficient than the multiphase MCMC only when n <\; ~100. This thesis
has three contributions. 1) It is analytically shown that\, unlike MCMC\,
the convergence of NSMC in the estimation of n-volume of a domain is not a
necessary condition for its convergence in any other integration over tha
t domain. 2) A direct numerical comparison of the naive NSMC and the mul
tiphase MCMC was performed for estimating n-volumes and different types of
integrands\, establishing this advantage in integration even over typical
convex domains when n <\; ~ 100. 3) A method for importance sampling is
suggested for NSMC with a demonstration of the improved performance in hi
gher dimensions for domains with heavy tailed extent density functions. In
identifying and ensuring a local volume of interest is sampled adequately
\, this method employs efficient sampling in high-D cones with a target di
stribution.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events
END:VEVENT
BEGIN:VEVENT
UID:30@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240111T110000
DTEND;TZID=Asia/Kolkata:20240111T120000
DTSTAMP:20240108T080126Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-hybrid-cds-11-januar
y-2024-an-arbitrary-lagrangian-eulerian-volume-of-fluid-method-for-free-su
rface-and-floating-body-dynamics-simulation/
SUMMARY:Ph.D: Thesis Defense: HYBRID: CDS: 11\, January 2024 "An Arbitrary
Lagrangian Eulerian Volume of fluid method for free surface and floating b
ody dynamics simulation."
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\nPh.D. Thesis D
efense (HYBRID)\n\n\n\nSpeaker : Mr. Bhatta Bhanu Teja\n\nS.R. Number :06-
02-00-10-12-12-1-09588\nTitle : "An Arbitrary Lagrangian Eulerian Volume o
f fluid method for free surface and floating body dynamics simulation."\nR
esearch Supervisor :Prof. Sashikumaar Ganesan\nDate &\; Time : January
11\, 2024 (Thursday)\, 11:00 AM\n\nVenue : The Thesis Défense will be hel
d on HYBRID Mode\n\n# 102 CDS Seminar Hall /MICROSOFT TEAMS\n\n\n\nPlease
click on the following link to join the Thesis Defense:\n\nMS Teams link\n
\nhttps://teams.microsoft.com/l/meetup-join/19%3ameeting_OGE0ZDhlZTMtNTJhY
y00NDg1LWJiY2EtNjk2ZWRjOWY1OWE3%40thread.v2/0?context=%7b%22Tid%22%3a%226f
15cd97-f6a7-41e3-b2c5-ad4193976476%22%2c%22Oid%22%3a%2283157d9d-c640-4c35-
856e-b22b4ab79e4e%22%7d\n\nABSTRACT\nThe floating body dynamics is treated
as a Fluid-Structure Interaction (FSI) problem. A FSI problem is where th
e forces from the fluid move/deform the interacting structure\, and the mo
vement of the structure\, in turn\, influences the dynamics of fluid flow
resulting in a coupled set of partial differential equations. The problem
is especially challenging owing to the time changing nature of the domain
and the presence of multiple interacting phases. These kind of time changi
ng domain and multi physics problems are primarily dealt with moving domai
n or fixed grid techniques. In the moving domain technique\, the governing
equations are posed in the so called Arbitrary Lagrangian Eulerian (ALE)
formulation. In ALE formulation\, the interfaces are resolved by the mesh
and thus leads to very good mass conservation properties. But the method f
ails when there are large topological changes\, such as mixing and splitti
ng. This problem can be partly handled by fixed grid techniques where a sp
ecial function is used to represent various phases\, but the method has it
s drawbacks\, one of which is the interfaces cannot be represented precise
ly and smears with time which leads to mass conservation problems and ofte
n much finer mesh is needed to localize the interface. Also\, because of t
he pure convection nature of the phase transport equation\, a naive/standa
rd discretization results in undershoots and overshoots of the solution. S
pecial stabilization schemes have to be used to suppress the oscillations.
The aim of the thesis is to treat the floating structure as a rigid body
to estimate its overall stability in free surface flows.\n\nFirst\, the pr
oblem is posed in moving mesh or Arbitrary Lagrangian Eulerian framework.
The motion of the free surface was captured. But the method failed to capt
ure the dynamics of the floating structure when it is introduced. Fine tun
ing the mesh yielded only incremental result. So the research focus was sh
ifted to fixed grid techniques\, particularly the 'Volume of Fluid' (VoF)
method. For the present problem\, we took a hybrid approach. The interface
between the floating structure and surrounding fluid/s is treated in a La
grangian way\, thus necessitating mesh movement\, and the fluid-fluid inte
rface (in our case\, it can be considered as water-air) is captured by the
VoF equation. As there is mesh movement\, the VoF equation was also posed
in ALE form.\n\nAs the VoF equation is a pure convection equation\, a nai
ve Galerkin discretization results in undershoots and overshoots in the so
lution. The Streamlined Upwind Petrov Galerkin (SUPG) stabilization is use
d to stabilize the VoF equation. The scheme is shown to give stable result
s. The Finite element method is used to discretize the coupled set of part
ial differential equations. The method is extensively discussed in a movin
g mesh setting with various boundary conditions. The partitioned time step
ping is used to march in time across phases\, and a fully implicit scheme
is employed within each phase.\n\nFinally\, this hybrid ALE-NSE-VoF with S
UPG stabilization scheme is shown to give stable results for extended time
steps. The numerical results with both the formulations(ALE and VoF) are
discussed. The simulations are carried out in distributed setting with Mes
sage passing interface(MPI)\, and the speedup results are discussed as wel
l.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VEVENT
UID:27@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240105T150000
DTEND;TZID=Asia/Kolkata:20240105T160000
DTSTAMP:20240103T050523Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-05th-january-2024-multim
odality-for-multilingual-nlp-the-need-an-application-and-open-questions/
SUMMARY:{Seminar} @ CDS: #102 : 05th January\, 2024 : "Multimodality for Mu
ltilingual NLP: The need\, an application\, and open questions"
DESCRIPTION:Department of Computational and Data Sciences\n\nDepartment Sem
inar\n\n\n\nSpeaker : Ms. Simran Khanuja\nTitle : "Multimodality for Multi
lingual NLP: The need\, an application\, and open questions "\nDate &\;
Time : January 05\, 2024\, 03:00 PM\n\nVenue : # 102\, CDS Seminar Hall\n
\n\n\nABSTRACT\n\nOur world is multimodal and multilingual\, demanding NLP
technology to be reflective of the same. From a cognitive perspective\, m
ultimodality is intricately linked with multilingualism in the brain. From
an application perspective\, this would make models capable of interpreti
ng and generating human-like responses. In this talk\, we will first prese
nt a bird's eye view on the benefits of multimodal modeling for multilingu
al NLP (the need). We will specifically focus on the visual and text modal
ity\, and discuss how real-world tasks would benefit from such models. Nex
t\, we will narrow down on our focus to one such application\, that of tra
nslating images across cultures (an application). Specifically\, we will d
iscuss our recent work on assessing the state of generative AI to do the t
ask\, and highlight existing gaps and biases of these models. Finally\, we
will broaden the discussion to explore open research questions in this ex
citing field\, and consider potential strategies for addressing these issu
es (open questions).\n\nBIOGRAPHY\n\nSimran Khanuja is a second-year PhD s
tudent at Carnegie Mellon University\, working with Prof. Graham Neubig. H
er research interests lie in multilingual\, multimodal modeling\, and down
stream applications that would benefit from such models. Prior to this\, s
he worked at Google and Microsoft Research in India for three years\, cont
ributing to the development of models and benchmarks for Indian languages.
\n\nHost Faculty: Dr. Danish Pruthi\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:28@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240105T110000
DTEND;TZID=Asia/Kolkata:20240105T120000
DTSTAMP:20240103T050807Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-05th-january-2024-solvin
g-pdes-with-neural-operators/
SUMMARY:{Seminar} @ CDS: #102 : 05th January\, 2024 : "Solving PDEs with Ne
ural Operators"
DESCRIPTION:Department of Computational and Data Sciences\n\nDepartment Sem
inar\n\n\n\nSpeaker : Prof. Siddhartha Mishra\, Chair Professor at ETH Zur
ich\, Switzerland\nTitle : "Solving PDEs with Neural Operators "\nDate &am
p\; Time : January 05\, 2024\, 11:00 AM\n\nVenue : # 102\, CDS Seminar Hal
l\n\n\n\nABSTRACT\n\nPartial differential equations (PDEs) are ubiquitous
as mathematical models in Science and Engineering. Solutions to both forwa
rd and inverse problems for PDEs can be encapsulated in terms of the so-ca
lled solution operators\, i.e\, mapping between infinite-dimensional funct
ion spaces\, that map inputs such as coefficients\, sources\, initial and
boundary conditions to the PDE solution. Learning the solution operator fr
om data falls under the rubric of "operator learning"\, a rapidly evolving
field within machine learning. In contrast to standard deep learning\, th
e inputs and outputs in operator learning are infinite-dimensional. Hence\
, special attention needs to be paid to the correspondence between the con
tinuous operator and its discrete realizations. By expanding on notions of
continuous-discrete equivalence in signal processing and harmonic analysi
s\, we introduce Representation equivalent Neural Operators (ReNOs) and sh
ow how they are a suitable framework of structure preserving operator lear
ning. Moreover\, a concrete instantiation of ReNOs\, the convolutional neu
ral operator (CNO) is presented and demonstrated as the state of the art m
achine learning surrogate for a wide variety of PDE benchmarks. If time pe
rmits\, we will discuss further applications such as learning PDE inverse
problems with Neural Inverse Operators.\n\nBIOGRAPHY\n\nSiddhartha Mishra
is a Chair Professor at ETH Zurich\, Switzerland\, where he heads the Comp
utational and Applied Mathematics Laboratory (CamLab)\, within the Seminar
for Applied Mathematics at the Department of Mathematics. He is also the
director of Computational Science Zurich and a core faculty member of the
ETH AI center. Mishra's research interests are in numerical analysis of PD
Es\, scientific computing and machine learning and in applications to flui
d dynamics\, geophysics\, astrophysics\, climate science and engineering.
For his contributions to research\, Mishra has received many awards and ho
nors\, including the Collatz Prize of ICIAM (2019)\, the Germund Dahlquist
Prize of SIAM (2021)\, the Rossler Prize of ETH (2023) and the Infosys Pr
ize (2019). Mishra has also been a keynote speaker at leading internationa
l conferences such as the International Conference of Mathematicians (ICM)
in 2018.\n\nHost Faculty: Prof. Sashikumaar Ganesan\n\n\n\nALL ARE WELCOM
E
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:26@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20240104T150000
DTEND;TZID=Asia/Kolkata:20240104T160000
DTSTAMP:20240102T055638Z
URL:https://cds.iisc.ac.in/events/cloud-seminar-pattern-aware-graph-mining
-abstractions-performance/
SUMMARY:[CLOUD SEMINAR] Pattern-Aware Graph Mining: Abstractions & Performa
nce
DESCRIPTION:\n\nCLOUD COMPUTING SEMINAR SERIES\n\n\n\nTITLE: Pattern-Aware
Graph Mining: Abstractions &\; Performance\n\nSPEAKER: Prof. Keval Vora
\, School of Computing Science\, Simon Fraser University\, Canada\n\nDATE/
TIME: Thu Jan 4\, 3PM\n\nVENUE: CDS Seminar Room #102\n\n\n\nABSTRACT\n\nM
odern graph mining applications like Motif Counting and Frequent Subgraph
Mining analyze the structural properties of graphs (i.e.\, rely on the sub
graph isomorphism problem). These applications are not only computationall
y expensive\, but are also difficult to express due to the complexities in
volved in the nuanced structural constraints to satisfy the application re
quirements.\n\nIn this talk\, I will give a flavor of the challenges invol
ved and our recent efforts in developing scalable graph mining systems. I
will present our pattern-aware processing philosophy that enables easier e
xpression of complex graph mining use cases\, enabling efficient pattern m
atching strategies for fast subgraph exploration. Specifically\, I will di
scuss the Anti-Vertex construct to easily express neighborhood constraints
in subgraph queries. Then\, I will present Subgraph Morphing\, a general
technique that exploits structural similarities across different patterns
to accelerate graph mining systems.\n\nBIO\n\nKeval Vora is an Associate P
rofessor at the School of Computing Science at Simon Fraser University. He
received his Ph.D. from the Department of Computer Science and Engineerin
g at the University of California\, Riverside where he was advised by Prof
. Rajiv Gupta. He was also a visiting researcher at the University of Cali
fornia\, Irvine where he worked with Prof. Harry Xu. His research addresse
s challenges in building scalable modern data analytics systems\, with a f
ocus on graph data processing and management. His work lies at the interse
ction of runtime systems (often touching various parts of the technology s
tack) and algorithmic semantics (to build smarter solutions). He specializ
es in developing efficient techniques with provable guarantees for large-s
cale graph systems.\n\nHost: Yogesh Simmhan\n\nAbout: The IBM-IISc Hybrid
Cloud Lab (IIHCL) hosted at IISc is curating the Cloud Computing Seminar s
eries with guest speakers from Industry and Academia speaking about the la
test technologies and research on Cloud and edge computing\, distributed c
omputing systems\, and AI/ML/Big Data platforms.\n\n\n\nALL ARE WELCOME
CATEGORIES:Talks
END:VEVENT
BEGIN:VEVENT
UID:25@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20231222T110000
DTEND;TZID=Asia/Kolkata:20231222T120000
DTSTAMP:20231222T073909Z
URL:https://cds.iisc.ac.in/events/seminar-cds-102-22nd-december-matchings-
in-big-graphs-approximation-and-streaming-algorithms/
SUMMARY:{Seminar} @ CDS: #102 : 22nd December: "Matchings in Big Graphs: Ap
proximation and Streaming Algorithms"
DESCRIPTION:Department of Computational and Data Sciences\n\nDepartment Sem
inar\n\n\n\nSpeaker : Prof. Alex Pothen\, Purdue University\n\nTitle : "Ma
tchings in Big Graphs: Approximation and Streaming Algorithms"\n\nDate &am
p\; Time : December 22\, 2023\, 11:00 AM\n\nVenue : # 102\, CDS Seminar Ha
ll\n\n\n\nABSTRACT\n\nMatchings in graphs are classical problems in combin
atorial optimization and computer science\, significant due to their theor
etical importance and relevance to applications. Polynomial time algorithm
s for several variant matching problems with linear objective functions ha
ve been known for fifty years. However\, these algorithms fail to compute
matchings in big graphs with billions of edges. They are also not concurre
nt and thus practical parallel algorithms are not known.\n\nThis has led t
o work in the last twenty years on designing approximation algorithms for
variant matching problems with near-linear time complexity in the size of
the graphs. Approximation has thus become a useful paradigm for designing
parallel matching algorithms. In this talk I will report on fast approxima
tion algorithms and streaming algorithms for the maximization versions of
edge-weighted matching\, edge-weighted b-matching\, and the maximum k-disj
oint weighted matching problems. We will also describe applications to air
craft design\, traffic routing in data centers and load balancing in quant
um chemistry.\n\nBIOGRAPHY\n\nAlex Pothen is a professor of computer scien
ce at Purdue University. His research interests are in combinatorial scien
tific computing\, graph algorithms and parallel computing. He received the
George Polya prize in applied combinatorics from the Society for Industri
al and Applied Mathematics (SIAM) in 2021 for his work on graph coloring a
lgorithms to enable Jacobian and Hessian matrix computations for optimizat
ion. He is a Fellow of SIAM\, ACM and AMS.\n\nHost Faculty: Prof. Sashikum
aar Ganesan\n\n\n\nALL ARE WELCOME
CATEGORIES:Events,Talks
END:VEVENT
BEGIN:VEVENT
UID:24@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20231221T110000
DTEND;TZID=Asia/Kolkata:20231221T120000
DTSTAMP:20231215T045612Z
URL:https://cds.iisc.ac.in/events/m-tech-research-thesis-colloquium-cds-in
telligent-methods-for-cloud-workload-orchestration/
SUMMARY:M.Tech Research Thesis {Colloquium}: CDS : "Intelligent Methods for
Cloud Workload Orchestration."
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nM.Tech Research
Thesis Colloquium\n\n\n\nSpeaker : Mr. Prathamesh Saraf Vinayak\n\nS.R. Nu
mber : 06-18-01-10-22-21-1-19717\n\nTitle :"Intelligent Methods for Cloud
Workload Orchestration"\nResearch Supervisor: Dr. Lakshmi J\nDate &\; T
ime : December 21\, 2023 (Thursday) at 11:00 AM\nVenue : # 102 CDS Seminar
Hall\n\n\n\nABSTRACT\n\nCloud workload orchestration is pivotal in optimi
zing the performance\, resource utilization\, and cost-effectiveness of ap
plications in data centers. As modern businesses and IT operations are mig
rating their businesses to the cloud\, understanding the dynamics of cloud
data centers has become indispensable. Often\, two perspectives play a pi
votal role in workload orchestration in data centers. One is from the clou
d provider side\, whose goal is to provision as many applications as possi
ble on the available resources\, biding to SLA constraints and increasing
return on investment. Others are from the side of enterprises and individu
al customers\, often referred to as end users\, whose primary objective is
to ensure application performance with a reduced deployment cost. Contain
erization has gained popularity for deploying applications on public cloud
s\, where large enterprises manage numerous applications through thousands
of containers placed onto Virtual Machines (VMs). While the need for cost
-efficient placement in cloud data centers is undeniable\, the complexitie
s involved in achieving this goal cannot be understated. This problem is u
sually modeled as a multi-dimensional Vector Bin-packing Problem (VBP). So
lving VBP optimally is NP-hard and practical solutions requiring real-time
decisions use heuristics. This work explores the landscape of cloud data
centers\, emphasizing the significance of efficient bin packing in achievi
ng optimal cost and resource utilization. Traditional methods\, including
heuristics and optimal algorithms\, face limitations in handling continuou
s request arrivals and the dynamic nature of cloud workloads. Integer Line
ar Programming (ILP)\, which can provide optimal solutions for small probl
em sizes with tens of requests\, may take minutes to hours to complete\, e
ven at such scales. Moreover\, optimal algorithms inherently demand perfec
t knowledge of all current and future requests to be placed within the bin
s\, rendering them unsuitable for the dynamic and often unpredictable onli
ne placement scenarios prevalent in cloud setups. To address these challen
ges\, this work introduces a novel approach to solving VBP through Reinfor
cement Learning (RL)\, trained on the historical container workload trace
for an enterprise\, a.k.a CARL (Cost-optimized container placement using A
dversarial Reinforcement Learning). The proposed work evaluates the effect
iveness of CARL in comparison to traditional methods. CARL leverages histo
rical container workload traces\, learning from a semi-optimal VBP solver
while optimizing VM costs. The contributions of this research extend beyon
d traditional methods\, providing insights into the advantages and disadva
ntages of heuristics\, optimal algorithms\, and learning approaches. We tr
ained and evaluated CARL on workloads derived from realistic traces from G
oogle Cloud and Alibaba for placing 10\,000 container requests onto over 8
000 VMs. CARL is fast\, making placement decisions for request sets with 1
24 containers per second within 65ms onto 1000s of potential VMs. It is al
so efficient\, achieving up to 13.98% lower VM costs than baseline heurist
ics for larger traces. To push the boundaries further\, we use the Mixture
of Experts (MoE) strategy in CARL\, wherein we use multiple experts who h
elp CARL learn placement policies of various approaches combined. Includin
g an MoE strategy enhances CARL's adaptability to changes in workload dist
ribution\, ensuring competitive performance in scenarios with skewed resou
rce needs or inter-arrival times.\n\n\n\nALL ARE WELCOME
CATEGORIES:Events
END:VEVENT
BEGIN:VTIMEZONE
TZID:Asia/Kolkata
X-LIC-LOCATION:Asia/Kolkata
BEGIN:STANDARD
DTSTART:20221221T110000
TZOFFSETFROM:+0530
TZOFFSETTO:+0530
TZNAME:IST
END:STANDARD
END:VTIMEZONE
END:VCALENDAR