Seminars

View all Seminars  |  Download ICal for this event

Explorations in Interpretability, Model Composability and Robust Explainability

Series: Ph.D. Colloquium

Speaker: Mariamma Antony, Ph.D (Engg.) student, Dept. of CSA

Date/Time: Nov 02 10:00:00

Location: CSA Auditorium, (Room No. 104, Ground Floor)

Faculty Advisor: Prof. Chiranjib Bhattacharyya

Abstract:
Deploying deep learning models in the real world exposes a persistent gap between the conditions a model is trained under and the conditions it must operate in. The gap widens as large language models (LLMs) and vision-language models (VLMs) become the default building blocks, because retraining them for every new need or perturbation is prohibitively expensive. This talk studies four properties that decide whether such models can be trusted and maintained: interpretability by construction, editing behavior after training without retraining, robustness to real-world perturbation, and explanations that stay reliable under that perturbation. The work has two parts: general-purpose methods with theoretical guarantees, then a high-stakes deployment case study. Part I: Robust and explainable AI for healthcare deployment. Our case study is AI-aided diagnosis of chest X-rays sent through messaging apps such as WhatsApp. This is a common setting in resource-constrained healthcare, and it was the basis of XraySetu, a service we built during COVID-19. We identify and quantify two previously unreported failure modes in state-of-the-art networks, including Vision Transformers: prediction instability and out-of-lung saliency. We propose COVIDMT, a multi-task learning framework that improves robustness by 40% and explainability by 35% over single-task baselines. We then ask why multi-task learning helps, and when auxiliary tasks hurt. We define the PEER score, which measures task relatedness by how explanations co-vary under perturbation. Through a first-order analysis along an auxiliary-weight homotopy, we prove the PEER-deletion principle. It links the sign of PEER to the change in explanation instability when an auxiliary task is removed. Building on it, we prove that retaining exactly the auxiliary tasks with negative PEER (the KEEP rule) reduces explanation instability, under mild regularity assumptions. Existing multi-task methods must compute and store per-task gradients, which is infeasible for billion-parameter LLMs and VLMs. PEER is computed on a small surrogate model, so KEEP scales to them. Applying it, we build LLaVA-PEER, a vision-language model for chest X-ray diagnosis and report generation. It reduces explanation instability by up to 25%, and it degrades far less than existing medical VLMs when images are compressed. Finally, we extend the study to larger scale and to object detection. We identify a third failure mode, localization instability, and show that the mitigation carries over, improving prediction stability, explainability and localization by up to 80% across model families. Part II: Interpretability and composability. Rule-based models are interpretable by construction, but existing learners optimize accuracy at the expense of coverage and parsimony. We formalize CDPR (high-coverage, discriminative, parsimonious rule sets) as a constrained submodular maximization problem whose overlap constraint is neither a matroid nor a knapsack constraint. GraphRules solves CDPR directly through an overlap-graph formulation. We prove an approximation guarantee that recovers the classical (1 �?? 1/e) greedy bound when the graph is disconnected. For a relaxed variant that penalizes mean overlap, we show the objective is non-monotone submodular. A Greedy algorithm then gives a 1/4-approximation in nearly linear time. Empirically, the learned rules are more accurate and interpretable than existing methods, with more than a 2.5-fold improvement in average coverage. Next, we turn to task arithmetic, which edits large pre-trained models, including LLMs and VLMs, by adding or negating task vectors. We show the challenges in existing task arithmetic, introduce the Task Localization Score and prove that any deviation of the score from its ideal value certifies a failure of weight disentanglement. We propose TaFaL, which optimizes layer-wise feature reconstruction and has a closed-form solution. We prove that reconstruction error controls the change in task loss. The loss on added tasks is bounded, and the loss on negated tasks increases by a strictly positive, quantifiable amount. At zero reconstruction error, the score reaches its ideal values and weight disentanglement holds. TaFaL is training-free and improves over state-of-the-art baselines by more than 17% on vision, language and multi-modal tasks. It extends to LoRA modules, including those of a vision-language model, and it aligns LLMs with human preferences using only binary desirable/undesirable signals, with no reward model and no reinforcement learning. It also reduces toxicity without fine-tuning. Together, these results treat interpretability, composability, robustness and explainability as connected requirements for trustworthy machine learning, including for LLMs and VLMs.\n\nMeeting Link :\nhttps://teams.microsoft.com/meet/467006301096861?p=kjzAPTxITHH65WZZFd