M.Tech. Projects Jan 2015

[M.Tech. Project Topics] | [Project Topics from Jan, 2016] | [Project Topics from Jan, 2014]

Proposed M.Tech. Computational Science Project Proposals

Jan 2015

  1. Unraveling Neural Computations Underlying cognition (Dr. Partha Talukdar and Dr. Sridharan Devarajan (CNS))The project involves applying novel dimensionality reduction, machine learning and multivariate analysis techniques to understand principles of neural computation in large-scale neural (EEG and fMRI) datasets. An additional sub-project includes developing distributed computing approaches for analysis of large neural datasets, for instance, based on the Map-Reduce approach.
  2. Deep Learning for identifying Neural Signatures of Cognitive States (Dr. Venkatesh Babu and Dr. Sridharan Devarajan (CNS))This project seeks to extend and apply the deep learning architecture to analyze spatiotemporal structure in brain activity recorded with human electroencephalography (EEG). The key challenge is to be able to identify cognitive states (e.g., whether a subject recognizes a visual scene, or to what object he/she is paying attention in a complex visual scene) based on recordings of brain activity.
  3. Convolutional Neural Networks for inferring the Neural Basis of Attention Control(Dr. Venkatesh Babu and Dr. Sridharan Devarajan (CNS))Natural visual scenes are complex and comprise multiple objects. Our brain deals with this information overload by paying attention to only one part of the scene at any given time, and shifting attention sequentially among various salient objects in the scene. This project seeks to use CNNs to model the process of allocating and shifting spatial attention in complex, natural visual scenes and to understand the brain regions associated with such attentional control.
  4. Goal-directed Knowledge Graph Expansion (Dr. Partha Talukdar)This project will explore expanding a Knowledge Graph in an entity-centric manner. For example, given an entity we would like to discover as many relevant information as possible. Prior experience with machine learning, text processing will be helpful, although not mandatory.
  5. Representation learning for Knowledge Harvesting (Dr. Partha Talukdar)This project will explore building on recent advances in representation learning and explore their applications for problems in knowledge harvesting. Prior experience with machine learning will be helpful.
  6. Efficient algorithms to compute the dual complex and the alpha complex (Prof. Vijay Natarajan, Prof. Sathish Vadhiyar)The dual complex of a set of weighted points is a triangulation of the space occupied by the points. The weight of a point may be interpreted as the radius of a ball centered at this point. The dual complex serves as an efficient data structure for representing the collection of balls and their adjacencies. In this project, we will study efficient algorithms for computing the dual complex and a particular subset called the alpha complex, which represents the volume occupied by the set of balls. We will explore parallel algorithms both in shared memory and distributed architectures.
    Fast algorithms for computing the dual complex and the alpha complex will particularly be important for applications to structural biology. We will also study the applications of the developed algorithm to speedup the computation and visualization of channels, cavities, and other structures in biomolecules.
  7. Supporting scalability with QoS on Container based cloud (Dr. J Lakshmi, Prof. SK Nandy)Most application developers find containers based clouds to be useful since they are lightweight but still provide the required software isolation. However, while consolidating such workload deployments on container based clouds the concern is low resource utilization. This project aims at exploring and evolving QoS constructs for enabling higher consolidation without loss of performance and scalability in applications.
  8. Exploiting hardware virtualization platforms for service level management and scalability in IaaS Clouds (Dr. J Lakshmi, Prof. SK Nandy) Clouds depend on system virtualization techniques for delivering better consolidation. Current trend in virtualized servers is to push resource sharing constructs to hardware for delivering fine-grained control and better performance. In this project we propose to leverage mechanisms provided by OpenPower platform for managing QoS. We aim to develop common standards based framework to manage QoS on IaaS/Cloud environment while allowing individual platform differentiations to be plugged in to the framework. Potentially, target OpenStack as IaaS platform of choice for proposing QoS management framework.
  9. Data provenance in cloud (Dr J Lakshmi, Prof SK Nandy) With compute clouds gaining wider adoption, many e-governance projects are exploring options for hosting their services in cloud. However, issues associated with physical location of data and related to its access and misuse draw attention to provenance related concerns. In this project we aim to explore the possibility of data tagging with its geo-spatial location and build a framework that addresses it provenance requirement.
  10. Performance modelling of VM migraion costs (Dr J Lakshmi, Prof. SK Nandy)VM migration is a convenient technique that has been claimed to be the most useful method for enabling applications to exploit elasticity in the cloud. Vm migration is a resource intensive operation and its expense is highly dependent on the application workload hosted in the VM. In this project we aim to build a performance model that can assess a given VM’s cost for migration.
  11. Sketch image analysis via Convolutional Neural Networks (CNN) (Dr. Venkatesh Babu)Convolutional Neural Networks (CNN) are biologically-inspired variants of Multi-Layer Perceptrons. This project aims to apply CNNs on the sketch images for the applications like classification and retrieval of the natural images based on the query sketches. The main challenge involved in the project is harnessing the sparsity of sketch images.
  12. Deep Learning for Action recognition (Dr. Venkatesh Babu)Deep Learning is a machine learning tool to characterize and analyze data at varioius levels of abstraction. This project aims to extend the deep learning architecture to videos for classifying the actions. Key challenge is to embed the temporal information in the deep architecture.
  13. Video Summarization (Dr. Venkatesh Babu) With the rapid growth in user-generated videos, it is becoming crucial to navigate them quickly and efficiently. This project aims to develop algorithms that automotically consolidate maximum information in a few frames either by extracting a set of keyframes(static video summary) from the original video or by summarizing the video shots in a more dynamic manner.
  14. Image Parsing (Dr. Venkatesh Babu) Image parsing attempts to find a semantically meaningful label for every pixel in an image. Depending on the labelling a task requires, the parsing problem is called as Image segmentation, Perceptual grouping or Object recognition.
  15. Transfer learning for Crowd behavior analysis (Dr. Venkatesh Babu)Crowd analysis is one of the key aspects of video surveillance. Learning amodel in one scene and transferring that knowledge to another scene canhelp in better prediction of the crowd behavior, avoiding the tediouslearning process for every situation.
  16. Blind Image Quality Assessment (Dr. Venkatesh Babu) Image quality assessment has been recently gaining attention due to thegrowing number of cameras and smart phones. The quality of the imagecaptured depends upon various factors including the lens properties,sensor noise, blur and compression artifact. This project aims to developalgorithm that can estimate the quality of the image without any referenceoriginal image.
  17. GPU based image super-resolution (Dr. Venkatesh Babu)The central aim of Super Resolution (SR) is to generate a higher resolution image from lower resolution images. High resolution image offers a high pixel density and thereby more details about the original scene. This project aims to efficiently parallelise the super-resolution algorithms utilizing GPU for speedup.
  18. Algebraic aspects of matching/grouping algorithms for graphs (Dr. Murugesan Venkatapathi) Methods for quantifying similarities in graphs and their partitioning will be explored. Relevant methods will be studied using matrix algebra and potential new parameters may be derived. Examples such as biological networks may be tested on. A student interested in matrix algebra and graphs/data-representation is preferred.
  19. Computational methods and Mean-field theories (Dr. Murugesan Venkatapathi) Computational solutions to large many-body problems using mean-field theories will be surveyed. Specific methods such as the dynamical mean-field theory with local density approximations to solve quantum many-body problems may be studied in more detail. Though the computational methods are the focus, a quick progressive understanding of the theoretical background is also needed for this exercise. Hence a student with strong aptitude in Physics and interest in computational methods is preferred.
  20. Job and Workflow Scheduling on OpenStack Private Clouds (Dr. Yogesh Simmhan)This project will explore scheduling algorithms and techniques for novel application workloads with dynamic characteristics that operate on elastic public Cloud infrastructure with varying VM performance. OpenStack private Cloud fabric will be used for benchmarking and validation.
  21. Complex Event and Stream Processing Across Edge and Cloud (Dr. Yogesh Simmhan)High velocity data from emerging applications like Internet of Things needs to be processed in a distributed manner, often across edge devices like smart phones and on Cloud data centres. This project will examine distributed Big Data platforms for stream and complex event processing, using sustainable water management as a usecase.
  22. Distributed Graph Processing (Dr. Yogesh Simmhan) This project will investigate distributed algorithms for static, dynamic and timeseries graphs using the subgraph-centric GoFFish platform. This includes developing new algorithms, analyzing them, tuning them for distributed execution, and comparative benchmarking against Apache Giraph, etc. using public and private Clouds.
  23. Distributed Graph Querying (Dr. Yogesh Simmhan) The project will develop a distributed and scalable graph database layered on top of the subgraph-centric GoFFish Big Data platform, similar to Hive for Hadoop. Topics include graph query specification, indexing techniques, decomposing queries into graph processing jobs, and scheduling them.
  24. Signal Processing Based Analysis of Photoacoustic Imaging (Dr. Phaneendra K Yalavarthy) Photoacoustic imaging is a scalable biomedical imaging modality that has applications starting from skin imaging to breast cancer screening. The photoacoustic imaging system matrix can be built using impulse responses of each image pixel, making in linear time-invariant (LTI) system. This project aims at analysing these impulse responses from the signal processing principles and explores the possibility of arriving at a convolution based (fourier space) image reconstruction.
  25. Evaluation of Compressive Sensing Data Collection Strategies for Biomedical Imaging (Dr. Phaneendra K Yalavarthy) Compressive sensing has revolutionised the biomedical imaging in terms of speeding up the acquisition and lowering the instrumentation cost due to its requirement of having lower data compared to traditional imaging. The compressive sensing algorithms performance largely dependent on the incoherence of the available limited data and the standard of acquisition of this limited data is do random sampling. Recent developments have been focussing on lowering the algorithm complexity for biomedical image reconstruction, with greedy algorithms being the most sought after ones. This project will look at the data space and evolve strategies to effectively assess the data collection strategies without performing the image reconstruction step. These kind of evaluations are necessary in imaging the humans, as the true (or expected) image is unknown.
  26. Development of Computationally Efficient Strategies for Fluorescence Image Denoising (Dr. Phaneendra K Yalavarthy) Image denoising has been a classic inverse problem and a routine image processing technique that is deployed to improve the image resolution/features. The fluorescence images of cellular and sub-cellular structures obtained in the microscope suffer from blur associated with the system response and aberrations. This blur is also spatially varying, denoising such images in not only challenging but very critical to obtain critical information. This project will develop efficient strategies for denoising such images and also evaluate them both numerically and experimentally for knowing their better utility.
  27. Hybrid execution strategies for graph applications on GPU systems (Prof. Sathish Vadhiyar) This project will look at building strategies for simultaneous use of both CPU and GPU cores in a GPU-based parallel system for efficient executions of graph applications including BFS, SSSP, satifiability problem, etc. The project will also explore efficient data organization strategies for fast access, and effective programming models for these applications on the GPU systems.
  28. Scaling Molecular Dynamics application to a large-scale parallel system (Prof. Sathish Vadhiyar) Molecular dynamics is an interesting parallel application with non-uniform computations and communications that vary over time and space. Providing parallelization with good scalability is challenging for this application. This project will build strategies to provide large-scale execution involving many thousand of cores.
  29. Job management on Supercomputer systems (Prof. Sathish Vadhiyar)Supercomputers have batch queues to which parallel jobs with specific requirements are submitted. These queues are statically configured during system installation. The configuration will impact both fairness and system usage. This project will dynamically reconfigure the queue configuration, including the number of queues and the parameters of each queues (the processor size limit etc.) based on the user job demands.