TBD

Federated Learning (FL) enables collaborative model training without centralizing data, but its distributed nature introduces new challenges. In this talk, I will discuss two key aspects: robustness and incentives. Robustness focuses on ensuring reliable learning under data heterogeneity and adversarial updates. Incentive design addresses how to motivate self-interested or strategic clients to contribute high-quality data and computation truthfully. I will outline recent approaches to each and reflect on how integrating the two can lead to more stable, fair, and trustworthy federated systems.

Kernel Machines are a classical family of models in Machine Learning that overcome several limitations of Neural Networks. These models have regained popularity following some landmark results showing their equivalence to Neural Networks. We propose a state of the art algorithm - EigenPro - based on gradient descent in the RKHS. This algorithm is much faster and requires less memory compared to previous attempts, and enables training large scale Kernel Machines over large datasets.

The stochastic gradient descent (SGD) algorithm is used for parameter estimation, particularly for massive datasets and online learning. Inference in SGD has been a generally neglected problem and has only recently started to get some attention. I will first introduce SGD for relatively simple statistical models and explain the limiting behavior of Averaged SGD. Then, I will present a memory-reduced batch-means estimator of the limiting covariance matrix that is both consistent and amenable to finite-sample corrections. Further, I will discuss the practical usability of error covariance matrices for problems where SGD is relevant, and present ongoing challenges in this area.

While foundation models and LLMs unlock new and unprecedented AI capabilities, they come with a substantially increased risk of memorising, regurgitating, and leaking privacy-sensitive data. Differential privacy, now a well-established standard for privacy protection, provides a principled solution to prevent such leakage, but is often computationally expensive (for good performance). I'll present some of our work on developing efficient and scalable algorithms AI inference and fine-tuning to make differential privacy practical in the era of foundation models. First, I will describe our method for generating artificial text with LLMs that is statistically similar to real data while preserving privacy. Our algorithm reduces the computational overhead of differential privacy from roughly 100-1000x in prior work to about 4x, making deployment feasible at scale. Next, I will discuss fine-tuning with differential privacy, where we build on a recent approach that injects correlated Gaussian noise across stochastic gradient steps. Our variant reduces the time complexity from quadratic to nearly linear, while maintaining comparable accuracy and privacy guarantees. I will conclude with a brief outlook on ongoing and future directions.

TBD

TBD

Simulation plays a key role in systems research. This session will begin with a gentle introduction to Discrete-Event Simulation (DES), covering event-driven and cycle-based approaches and explain why parallelizing DES is essential yet non-trivial. I will then introduce SiTAR, a parallel simulation framework that we have developed over several years, outline its key ideas, modeling language, and runtime, and share result on modeling and parallel simulation of multicore and memory subsystem models for design exploration. This talk will conclude with an overview of our ongoing work on automated model generation and hybrid (discrete-continuous) simulation.

The increasing complexity of integrated circuit (IC) design has made traditional Electronic Design Automation (EDA) approaches computationally intensive and time-consuming. Recent advances in Machine Learning (ML) are transforming the EDA landscape by enabling data-driven optimization, faster design closure, and improved accuracy in prediction-based tasks. This talk will provide an overview of how ML techniques are being integrated into various stages of the VLSI design flow-such as synthesis, placement, routing, and verification. Emphasis will be placed on key challenges, successful use cases, and future research directions where ML can significantly enhance design productivity and innovation in EDA.

TBD

TBD

Generative AI, especially Large Language Models (LLMs) and Multimodal Language Models (MLMs), is creating exciting opportunities in healthcare. However, real-world use is still challenging due to the need for models that are compact, personalized, safe, and capable of handling multiple languages and data types. This work tackles these challenges in three main directions: building specialized models, developing advanced methods for key healthcare tasks, and creating strong evaluation benchmarks. First, we build a small, domain-specific language model for veterinary medicine, a field that is often overlooked. This model is trained from scratch with proper pretraining, fine-tuning, and safety alignment. Second, we design models for summarizing medical inputs that include both text and images, focusing on low-resource and code-mixed languages to help healthcare professionals better understand complex patient data. Finally, we introduce new benchmarks to evaluate model performance in medical settings, including a) M3Retrieve, a large multimodal retrieval benchmark across 5 domains and 16 medical fields, and b) a multilingual trust benchmark that covers 15 languages and 18 detailed tasks. Together, these efforts aim to make generative AI more practical, reliable, and inclusive for healthcare use.

Sequential hypothesis testing is a fundamental tool in statistics and machine learning. It enables decisions to be made from streaming data while controlling errors. But what is the minimum number of samples that you need to make such decisions with high confidence? In this talk, we will see two tight lower bounds on the stopping times of power-one sequential tests that guarantee small false positive rates and eventual detection under the alternative. These results extend the classical works (Wald, Farrell) to modern, fully nonparametric composite settings, using an information-theoretic quantity, which we call KL-inf, the minimal KL divergence between the null and alternative sets. We will also see sufficient conditions for designing tests that match these lower bounds. Given past work, these upper and lower bounds are unsurprising in their form; our main contribution is the generality in which they hold, for example, not requiring reference measures or compactness of the classes.

The term Large Action Model (LAM) is typically used in the context of specific agentic workflows leveraging Large Language Models (LLMs) designed to accomplish tasks by interacting with tools/APIs, device controls, web pages or user Interfaces. Since such actions are often local and frequent, there is a compelling case to consider Small Action Models (SAMs) to unlock on-device, private, low-power, low-latency agents. However, small models pose significant challenges since they are often less reliable. This talk discusses the challenges in building high-quality, reliable on-device action models and presents an approach to building SAMs that are capable of complex query orchestration

In this talk, I will talk about some hardware/software work my group has done in the area of stochastic computing based machine learning acceleration. I will talk about suitability of the SC to this workload, how to deal with its inherent approximate nature and briefly discuss few chip prototypes which we leverage both logic and in-memory implementations of SC-based accelerators for dense as well as a sparse compute.

Deploying deep neural networks (DNNs) on edge devices presents unique challenges due to limited computational resources, diverse hardware architectures, and dynamic runtime conditions. This talk presents a comprehensive overview of our research efforts aimed at making DNN inference on heterogeneous edge platforms more efficient, adaptive, and responsive to real-world constraints.

We increasingly entrust large parts of our daily lives to computer systems that are growing ever more complex. Developing scalable and trustworthy methods for designing, building, and verifying these systems is therefore crucial. In this talk, I will focus on automated synthesis a technique that uses formal specifications to automatically generate systems such as functions, programs, or circuits that provably satisfy their requirements. Can we leverage AI-particularly machine learning and large language models-to propose candidate programs or functions? Yes. But can these candidates be guaranteed correct? Can we verify them, or even repair them so that they provably meet the intended specifications? This talk will revolve around these questions-how AI and Automated Reasoning can be effectively fused to synthesize reliable systems.

Understanding why Graph Neural Networks (GNNs) make certain predictions remains a central challenge. Existing explainability methods, though insightful, often produce complex or large explanations that are difficult for humans to interpret, and they primarily focus on local reasoning around individual predictions. Yet, a GNN learns global reasoning patterns that govern its behavior across data. This motivates our broader vision-to design explainability algorithms that are both faithful to the model's reasoning and interpretable to humans. We first addressed this through GraphTrail, the first global, post-hoc GNN explainer that represents model behavior as Boolean formulae over subgraph-level concepts discovered using Shapley values, offering a symbolic understanding of GNN reasoning. However, graph datasets are inherently multi-modal, combining topology with rich node and edge attributes, and a graph with n nodes admits up to 2n subgraphs, making the isolation of neural reasoning patterns combinatorially prohibitive. Consequently, GraphTrail is limited to small, labeled graphs. Our latest work, GNNXemplar, overcomes these challenges. Inspired by Exemplar Theory in cognitive science, it identifies representative nodes-exemplars-in the embedding space and derives interpretable natural language rules for their neighborhoods using large language models.

The rapid growth of edge devices has created a dynamic, AI-powered data ecosystem with significant potential for societal advancement. However, privacy concerns restrict data sharing across multiple owners, hindering the full potential of AI. Furthermore, edge devices often have substantial resource constraints and heterogeneity, severely restricting their ability to handle large models. This talk will provide innovative solutions to overcome these challenges and enable efficient and privacy-preserving ML in diverse edge settings. As a key highlight, we will address the following question: How to enable federated learning of a large global model, when each edge device can only train a small local model?

Artificial intelligence has not only transformed applications but has also redefined the design principles of modern computing systems. This presentation explores the emerging co-evolution of systems that drive AI and AI that shapes systems. It first discusses system-level innovations in AI accelerators, including dataflow architectures, memory hierarchies, and interconnect optimizations that enable scalable and energy-efficient model execution. In parallel, it examines the use of AI techniques to enhance system intelligence, with particular emphasis on learning-guided cache management and adaptive memory control. By integrating predictive models into runtime and architectural decisions, computing platforms can dynamically adapt to diverse workloads and access patterns, leading to more efficient and intelligent system behavior.

TBD

Serving Large Language Models (LLMs) at scale involves managing a wide range of workloads with varying performance requirements. Applications such as interactive chatbots demand low latency, while tasks like document generation can tolerate longer response times. This talk presents work from the M365 Research group on optimizing LLM inference infrastructure to efficiently balance these diverse needs. We introduce a dynamic framework that combines short-term request routing with long-term resource management, including GPU scaling and model placement, guided by optimization techniques and a fairness-aware scheduler.

Performance models assume steady state and infinite resources. Real systems don't. Learn how performance engineers diagnose where theory breaks, design experiments to expose the gaps, and use that understanding to solve hard problems. We'll explore the principles through real debugging examples.

Oceanographers struggle with the scalability that is required for visual analysis of massive and multivariate ocean model output for tasks like event tracking and phenomenon identification. Our research addresses this challenge by introducing a novel methodology centered on two key innovations: integrating specialized domain-specific analysis modules directly as efficient ParaView filters, and developing parallel solutions that leverage the use of computing resources available to the analyst. This approach culminated in the development of pyParaOcean, an extendible and easily deployable visualization system that leverages ParaView's parallel processing capabilities. We demonstrate the utility of this research with a Bay of Bengal case study and present scaling studies that confirm the high efficiency of the system.

TBD

Neural Radiance Fields (NeRFs) have offer remarkable scene reconstructions through differentiable volumetric rendering. Yet, their reliance on dense image capture and heavy optimization limits practical deployment. This talk will first introduce the foundations of NeRFs — their scene representation, rendering pipeline, and learning dynamics — and then discuss advances that extend NeRFs to sparse input settings. I will also introduce 3D Gaussian Splatting, a recent point-based alternative enabling real-time rendering, and highlight our recent work on sparse-input 3D Gaussian Splatting.

TBD