Relevant Thesis-Based Degree Programs
Graduate Student Supervision
Doctoral Student Supervision
Dissertations completed in 2010 or later are listed below. Please note that there is a 6-12 month delay to add the latest dissertations.
Budget-robust active learning (2025)
Deep learning has made significant strides in recent years, largely due to the availability of vast amounts of labeled data. However, expensive and time-consuming manual annotation limits the widespread adoption of Artificial Intelligence (AI), particularly for smaller organizations and individuals. This highlights the need for data-efficient AI frameworks that reduce dependence on large-labeled datasets, making AI more accessible. Active learning, where a model strategically selects the most informative data points for annotation, offers a promising solution to this challenge. It improves model performance with fewer labeled examples, making it especially valuable in domains where labeling is costly.Recent research has revealed that the effectiveness of active learning methods varies significantly across different budget regimes, where the budget is defined by the size of a labeled set. In particular, uncertainty-based methods, which perform well in high-budget settings, often underperform compared to representation-based methods or even random sampling in low-budget regimes. In this thesis, we investigate how to improve active learning under both high- and low-budget regimes. We begin with the high-budget setting, where we introduce a novel uncertainty-based method that leverages neural tangent kernels (NTKs) to make computation of look-ahead acquisition strategies feasible. This approach allows the model to account for the changes of ``future" predictions, resulting in strong performance across various datasets, particularly in the high-budget regimes.In the low-budget regimes, we propose MaxHerding, a representation-based method that generalizes the recently introduced ProbCover and establishes connections to other low-budget active learning techniques. To further explore active learning under limited annotation budgets, we consider its application to meta-learning (or few-shot learning) and develop a simple yet effective acquisition strategy based on Gaussian Mixture Models (GMMs), motivated by a max-margin classifier.Given the difficulty of determining the appropriate budget regime in advance, we finally propose Uncertainty Herding (UHerding), a budget-robust active learning method that adaptively interpolates between uncertainty- and representation-based strategies. Our empirical results show that UHerding consistently outperforms existing methods across a wide range of budget regimes, offering a promising direction toward hyperparameter-free and more robust active learning in real-world applications.
View record
Learning dynamics of deep learning -- force analysis of deep neural networks (2025)
This thesis investigates the learning dynamics of deep learning systems through a local, physics-inspired analytical lens. Motivated by the need for fine-grained insights into model behavior, we begin with the step-wise influence that a single training example exerts on a specific observing example during learning. Central to our approach is the proposed AKG decomposition, which dissects this influence into three interpretable components: similarity (K), normalization (A), and prediction gap (G). This decomposition enables an analogy with classical force analysis: the force originates from G, is shaped by AK, and is ultimately applied to the target object, e.g., to the model confidence, output, hidden representations, or parameters. Building upon this foundation, we gradually scale the analysis from individual interactions to cumulative effects over time, akin to tracking an object’s motion under multiple forces. We apply it to the following problems. Supervised classification: We study the learning trajectories of examples with varying difficulty and reveal an interesting "zig-zag" pattern that emerges during optimization. Our analysis explains this behavior and inspires a novel knowledge distillation method, Filter-KD, which improves supervision signals for student models. Large language model (LLM) finetuning: We extend the framework to account for the autoregressive nature of LLMs and the presence of negative gradients. The unified perspective explains behaviors across finetuning methods such as SFT, DPO, and GRPO. We also highlight the critical role of negative gradients. In particular, we identify the "squeezing effect": a counterintuitive phenomenon caused by improperly applied gradient ascent. Representation learning: We explore the dynamics of hidden features, revealing how adaptation energy and directions influence the feature drift. Our analysis leads to a provable pattern of feature adaptation in a head-probing then finetuning pipeline, offering insights and inspiring several practical strategies. Simplicity bias and compositional learning: Revisiting foundational questions about why structured representations are learned faster, we apply our framework to a compositional learning setting. Our findings align with principles such as Occam’s Razor and the idea of "compression for AGI," offering a novel dynamical explanation rooted in compression and learning speed.
View record
Master's Student Supervision
Theses completed in 2010 or later are listed below. Please note that there is a 6-12 month delay to add the latest theses.
Insights from infinitely wide neural networks (2024)
Studying neural networks in the limit of infinite-width has provided us with numerous valuable theoretical and practical insights about the initialization of NNs, their training dynamics and properties of the learnt functions. One of the theoretical tools emerged from this study is the empirical Neural Tangent Kernel (eNTK). The eNTK can provide a good understanding of a given network’s representation: they are often far less expensive to compute and applicable more broadly than infinite-width NTKs. In this work, we use eNTKs to predict the local dynamics of neural networks in an active learning setup to propose a new method for approximating active learning acquisition strategies that are based on retraining with hypothetically-labeled candidate data points. Furthermore, to tackle the notorious space and computational complexity of calculating eNTKs, we propose a fast approximation of eNTK using a block-diagonal kernel resulting from eNTK with respect to only one (or average) of the output neurons of a network. We further use this approximation in our proposed “look-ahead” strategies in deep active learning. We finally present empirical evidence that our querying strategy beats other look-ahead strategies by large margins, and achieves equal or better performance compared to state-of-the-art methods on several benchmark datasets in pool-based active learning.
View record
A kernel-based approach to differentially private image generation (2023)
The gold standard privacy notion, differential privacy (DP), has gained widespread adoption in academic research, industry products, and government databases due to its mathematically provable privacy guarantee. However, the composability property of DP leads to privacy degradation with multiple accesses to the same data. Differentially private data generation has emerged as a solution, creating synthetic datasets resembling private data while allowing repeated access without additional privacy loss. Existing methods often assume specific use cases for synthetic data, limiting flexibility.This thesis addresses the challenge of producing flexible synthetic data by leveraging deep generative modeling and addressing privacy loss in other methods such as generative adversarial networks (GAN). we propose utilizing public data to learn perceptual features (PFs) for comparing real and synthetic data distributions, employing a non-adversarial generator training scheme based on Maximum Mean Discrepancy (MMD) to mitigate privacy loss.Experimental results reveal the efficacy of our method. it successfully generates samples for CIFAR-10, CelebA, MNIST, and FashionMNIST. Theoretical analysis of our privacy-preserving loss function clarifies the privacy-accuracy trade-offs.
View record
Kernel methods for invariant representation learning : enforcing fairness and conditional independence (2023)
This work introduces two novel kernel-based measures to enforce certain invariance properties in the learned representation space of a deep neural network. The first method, MMD-B-Fair, learns fair representations of data via kernel two-sample testing. It finds neural features of data where a maximum mean discrepancy (MMD) test cannot distinguish between different representations of different sensitive groups, while preserving information about the target variable to be predicted. To minimize the power of an MMD test this method exploits the simple asymptotics of a block testing scheme to address challenges presented by the complex dependency of the test threshold on the estimated MMD. Compared to existing methods on fair representation learning, MMD-B-Fair does not require generative modeling or discriminative architectural tuning, and is able to achieve competitive results on fairness benchmarks and downstream transfer. The second method, CIRCE, introduces a measure of conditional independence for multivariate continuous-valued variables that can be efficiently used as a regularizer to learn deep neural features that are conditionally independent of a known distractor Z given a target label Y. CIRCE requires just a single ridge regression from Y to kernelized features of Z, which can be done in advance. It is then only necessary to enforce independence of the learned neural features from the residuals of this regression. By contrast, earlier measures of conditional dependence require multiple regressions for each step of feature learning, resulting in severe bias and variance, and greater computational cost. CIRCE has superior performance to previous methods on challenging benchmarks, including learning conditionally invariant image features. Python implementations of both methods are made publicly available at github.com/namratadeka/mmd-b-fair and github.com/namratadeka/circe.
View record
If this is your researcher profile you can log in to the Faculty & Staff portal to update your details and provide recruitment preferences.
Membership Status
Program Affiliations
Academic Unit(s)