CS KathaBarta | ସଂଗଣକ ବିଜ୍ଞାନ କଥା ବାର୍ତା

Deep Learning models have now become ubiquitous. However, the hardware requirements to deploy SoTA models are increasing faster than Moore's law can deliver. Model compression, particularly pruning, is one way to alleviate this problem. This talk explores challenges with pruning in multi-branch architectures and how coupled channels (CCs) impact inference time. It also focuses on pruning without access to training data, a growing necessity due to privacy and computational constraints.

TBA

Multi-Agent Retrieval-Augmented Generation (RAG) enhances AI-driven reasoning by combining specialized agents for retrieval, analysis, and content generation. This collaborative approach improves accuracy, minimizes hallucinations, and scales efficiently across applications like research assistants, legal analysis, and customer support. This talk explores multi-agent RAG’s architecture, benefits over single-agent models, and key challenges, including agent coordination and retrieval bias. We discuss optimization strategies, reinforcement learning for self-improvement, and future research directions. Multi-agent RAG represents the next step in intelligent, fact-based AI reasoning, bridging structured retrieval with advanced generative capabilities.

Generative models have the potential to revolutionize biomedical imaging. However, in this mission-critical domain, the evaluation of generative models is imperative to ensure patient safety. Designing evaluation frameworks for generative models is essential at different stages in the biomedical-tool development process. In this talk, I will focus on generative models of images for data augmentation applications. I will discuss the limitations of some existing approaches and present some of my work on (i) rule-out tests, (ii) in-silico testing and (iii) post-hoc testing for clinical datasets.

I will give a high-level overview of graph representation learning that uses homomorphism counting in one way or the other. Homomorphism counting is a way to measure 'how often' a given pattern appears in a graph. In the context of (message passing) graph neural networks (GNNs), homomorphism counting is mainly used in two areas: (1) Quantifying what GNNs can do and where they fail and (2) improving the capabilities of GNNs. Special interest in both these areas lies on the expressivity of GNNs, i.e., their ability to learn different representations for nonisomorphic graphs. I will present and discuss the basics, as well as recent results in this interesting and active area of research.

Instance-level explanation of graph neural networks (GNNs) is a well-studied area. These explainers, however, only explain an instance (e.g., a graph) and fail to uncover the combinatorial reasoning learned by a GNN from the training data towards making its predictions. In this work, we introduce GraphTrail, the first end-to-end, global, post-hoc GNN explainer that translates the functioning of a black-box GNN model to a boolean formula over the (sub)graph level concepts without relying on local explainers. GraphTrail is unique in automatically mining the discriminative subgraph-level concepts using Shapley values. Subsequently, the GNN predictions are mapped to a human-interpretable boolean formula over these concepts through symbolic regression. Extensive experiments across diverse datasets and GNN architectures demonstrate significant improvement over existing global explainers in mapping GNN predictions to faithful logical formulae. The robust and accurate performance of GraphTrail makes it invaluable for improving GNNs and facilitates adoption in domains with strict transparency requirements.

Reference : Armgaan, Burouj, Manthan Dalmia, Sourav Medya, and Sayan Ranu. "GraphTrail: Translating GNN predictions into human-interpretable logical rules." Advances in Neural Information Processing Systems 37 (2024): 123443-123470.

The Materials Simulation team of BIOVIA has been engaged in the development of electronic and molecular simulation applications for over three decades. At the centre of this activity is the notion of multi-scale physics. An overview of three contemporary areas of interest spanning a number of scales will be discussed: the development of machine learnt potentials, in conjunction with the University of Cambridge1, for use in molecular dynamics simulations, the application of mesoscale methods to the simulation and understanding of lipid bilayer and liposome formation2 and the extension of the molecular dynamics method to granular dynamics3 with an application to battery electrode structure analysis.

1. Batatia, I. et al "MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields", Advances in Neural Information Processing Systems, 11423-11436 (2023).
2. Development of Innovative Drugs Using Materials Studio
3. Jing Bi et al “Multiscale modeling for the science and engineering of materials” International Journal for Multiscale Computational Engineering, 19, 1-80 (2021).

Electricity markets and their associated operational structures play a pivotal role in ensuring a secure, reliable, and economically efficient energy supply. By facilitating competition among diverse power producers - from conventional fossil-fuel generators to renewable energy sources - electricity markets promote transparency in pricing and stimulate innovation. Key market mechanisms, such as day-ahead auctions and intraday trading, help align generation with demand, encouraging optimal resource allocation and minimizing imbalances in real time.On an operational level, Transmission System Operators (TSOs) oversee grid reliability through balancing markets and ancillary services, ensuring electricity flows reliably across interconnected networks. These operators deploy mechanisms like reserve capacity activation and frequency control to stabilize the system whenever unexpected deviations in supply or demand occur. Furthermore, the integration of distributed energy resources and the rise of prosumers introduce new complexity and opportunities, including peer-to-peer trading and localized market initiatives. Overall, the interplay between electricity market design and operational frameworks underpins the modern power sector’s resilience, adaptability, and efficiency. By continuously refining these market structures and coordinating operations across borders and jurisdictions, stakeholders can foster a robust and sustainable energy ecosystem that meets evolving societal needs.