BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.2.3.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:100@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20250117T103000 DTEND;TZID=Asia/Kolkata:20250117T113000 DTSTAMP:20250110T121730Z URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-cds-january-17-2025- learning-from-limited-and-imperfect-data/ SUMMARY:Ph.D. Thesis Defense: CDS: January 17\, 2025 "Learning from Limited and Imperfect Data" DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\n\nPh.D. Thesis D efense\n\n\n\nSpeaker : Mr. Harsh Rangwani\nS.R. Number : 06-18-01-10-12-1 9-1-17477\nTitle : "Learning from Limited and Imperfect Data"\nResearch Su pervisor : Prof. Venkatesh Babu\nThesis examiner : Prof. Richa Singh\, IIT Jodhpur\nDate &\; Time : January 17\, 2025 (Friday) at 10:30 AM\nVenue The Thesis Defense will be held on HYBRID Mode\n# 102 CDS Seminar Hall /M ICROSOFT TEAMS.\n\nPlease click on the following link to join the Thesis D efense\n\nMS Teams link\n\n\n\nABSTRACT\n\nDeep Neural Networks have demon strated orders of magnitude improvement in capabilities over the\nyears af ter AlexNet won the ImageNet challenge in 2012. One of the major reasons f or this success is the availability of large-scale\, well-curated datasets . These datasets (e.g.\, ImageNet\, MSCOCO\, etc.) are often manually bala nced across categories (classes)\nto facilitate learning of all the catego ries. This curation process is often expensive and requires throwing away precious annotated data to balance the frequency across classes. This is b ecause the distribution of data in the world (e.g.\, internet\, etc.)\nsig nificantly differs from the well-curated datasets and is often over-popula ted with samples from common categories. The algorithms designed for well- curated datasets perform suboptimally when used to learn from imperfect d atasets with long-tailed imbalances\nand distribution shifts. For deep mod els to be widely used\, getting away with the costly curation process by d eveloping robust algorithms that can learn from real-world data distributi on is necessary. Toward this goal\, we develop practical algorithms for De ep\nNeural Networks that can learn from limited and imperfect data present in the real world. These works are divided into four segments\, each cove ring a scenario of learning from limited or imperfect data. The first part of the works focuses on Learning Generative\nModels for Long-Tail Data\, where we mitigate the mode-collapse for tail (minority) classes and enable diverse aesthetic image generations as head (majority) classes. In the se cond part\, we enable effective generalization on tail classes through Ind uctive Regularization\nschemes\, which allow tail classes to generalize as the head classes without enforcing explicit generation of images. In the third part\, we develop algorithms for Optimizing Relevant Metrics compare d to the average accuracy for learning from long-tailed data\nwith limited annotation (semi-supervised)\, followed by the fourth part\, which focuse s on the effective domain adaptation of the model to various domains with zero to very few labeled samples.\n\nGenerative Models for Long-Tail Data. We first evaluate generative models’ performance\, specifically variant s of Generative Adversarial Networks (GANs) on long-tailed datasets. The G AN variants suffer from either mode-collapse or miss-class modes during ge neration. To miti- gate this\, we propose Class Balancing GAN with a Class ifier in the Loop\, which uses a classifier to asses the modes in generate d images and regularizes GAN to produce all classes equally. To alleviate the dependence on the classifier\, following our observation that spectral norm explosion of Batch Norm parameters is the major reason for mode coll apse. We develop an inexpensive group Spectral Regularizer (gSR) to mitiga te the spectral collapse\, which significantly improves the SotA condition al GANs (SNGAN and BigGAN) performance on long-tailed data. However\, we o bserved that class confusion was present in the generated images due to no rm regularization. In our latest work NoisyTwins\, we factor the latent sp ace as distinct Gaussian by design for each class\, enforcing class consis tency and intra-class diversity using a contrastive approach (BarlowTwins) . This helps to scale high-resolution StyleGANs for ≥ 1000 class long-ta iled datasets of ImageNet-LT and iNaturalist2019\, achieving state-of-the- art (SotA) performance.\n\nInducting Regularization Schemes for Long-Taile d Data. While Data Generation is exciting for improving classification mod els on tail classes\, it often comes with the cost of training an auxiliar y GAN model. Hence\, a lightweight technique like enhancing loss weights ( re-weighting) for tail classes while training CNNs is practical to improve minority class performance. However\, despite this\, the model only attai ns minima for the head class loss and converges to saddle point for tail c lasses. We show that inducing inductive bias of escaping saddles and conve rging to minima for tail classes\, using Sharpness Aware Minimization (SAM ) significantly improves performance on tail classes. Further training Vis ion Transformer (ViT) for long-tail recognition is hard\, as they don’t have inductive biases like locality of features\, which makes them data hu ngry. We propose DeiT-LT\, which introduces OOD and low-rank distillation from CNN to induce CNN-like robustness into scalable ViTs for robust perfo rmance.\n\nSemi-Supervised Learning for Practical Non-Decomposable Objecti ves. 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 unlabele d data present (i.e. semi-supervised setting). For this\, we introduce a p aradigm where we measure the performance using relevant non-decomposable m etrics such as worst-case recall and recall H-mean on a held-out set\, and we use their feedback to learn in a semi-supervised long-tailed setting. We introduce Cost-Sensitive Self Training (CSST)\, which generalizes self- training based semi-supervised learning (e.g. FixMatch\, etc.) to the long -tail setting with strong guarantees and empirical performance. The genera l trend these days is to use self-supervised pre-training to obtain a robu st model and then fine-tune it. In this setup\, we introduce SelMix\, an i nexpensive fine-tuning technique to optimize the relevant metrics using pr e-trained models. In SelMix\, we relax the assumption that the unlabeled d istribution is similar to the labeled one\, making the models robust to di stribution shifts.\n\nEfficient Domain Adaptation. The long-tail learning algorithms focus on limited data setup and improving in-distribution gener alization. Still\, for practical usage\, the model must learn from imperfe ct data and perform well across various domains. Toward this goal\, we dev elop Submodular Subset Selection for Adversarial Domain Adaptation\, which carefully selects a few samples to be labeled for maximally improving mod el performance in the target domain. To further improve the efficiency of the Adaptation procedure\, we introduce Smooth Domain Adversarial Training (SDAT)\, which converges to generalizable smooth minima. The smooth minim um enables efficient and effective model adaptation across domains and tas ks\n\n\n\nALL ARE WELCOME CATEGORIES:Events,Thesis Defense END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20240118T103000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR