CNI CONNECT

Monthly Newsletter from CNI

Issue - February 2025

Upcoming Events

Professors, Sidharth Jaggi from University of Bristol and Nikhil Karamchandani from IIT Bombay, will be delivering their insightful talks on Group Testing on February 17th and 18th. We encourage you to attend and not miss this exciting opportunity!

Join for an engaging talk by Prof. Prasenjit Chanak from IIT BHU on 25th February 2025.

What’s cooking at CNI?

CNI PhD Scholar Srinivas's work on "The Power of Two in Large Service Marketplaces" has been accepted at IEEE INFOCOM 2025

CNI PhD Scholar Srinivas Nomula’s joint work with Dr. Dheeraj Narasimha, Prof. Srinivas Shakkottai and Prof. Parimal Parag on the effectiveness of selecting two servers uniformly at random, among all the available servers with independent pricing scheme has been accepted at IEEE INFOCOM 2025. Their work explores the design and optimization of large-scale service marketplaces, particularly in the context of cloud services. It addresses the challenge of efficiently matching incoming jobs to available servers while maximizing revenue and minimizing blocking probability.

The central argument is that a distributed marketplace model, where jobs are randomly matched to two servers which then competitively price their services, offers a scalable and efficient solution. This "power-of-two choices" approach outperforms single-server matching and approaches the performance of complex centralized systems, but with significantly reduced complexity. The paper uses mean-field game (MFG) theory to analyze the system's behavior, demonstrating the existence and convergence to a mean field equilibrium (MFE).

Data-Driven Agriculture to Sustainably Nourish the World

In this talk, Dr. Ranveer Chandra highlighted the role of data-driven agriculture in addressing global food security. Emphasizing sustainable farming amid climate challenges, he outlined Microsoft’s vision for a connected agricultural ecosystem, featuring AI-driven modeling, aerial imagery, and multimodal data integration. He stressed the need for affordable technology, citing innovations like TV white spaces for connectivity and smartphone-based soil assessments. Through case studies, he demonstrated these technologies’ impact on farm productivity and concluded with a call for collaboration in advancing sustainable, tech-driven agriculture.

PEARL, Personalized Exercise Agents using Reinforcement Learning

In the seminar, Narayan Hegde addressed global health challenges such as sedentary behavior, obesity, and metabolic syndrome, proposing personalized nudges as an effective strategy to promote physical activity. He emphasized the importance of optimizing content, timing, and frequency to sustain user engagement. Introducing PEARL, a system utilizing Large Language Models (LLMs) and Reinforcement Learning (RL) to personalize nudges based on walking patterns, he presented findings from large-scale studies showing RL's threefold improvement over traditional methods. The seminar concluded with an overview of key research initiatives at Google DeepMind India, highlighting recent advancements and future directions.

Exploring multi-bounce scattering to increase the field-of-view of mm-wave radar

Dr. Divyanshu Pandey’s talk explored the use of mm-wave radars to detect objects beyond their direct field of view, even when fully obstructed. Traditional radars treat multipath signals—such as bounces off walls or people—as noise. This talk introduced a framework that leverages these signals to perceive occluded objects, overcoming limitations of existing methods that require specific assumptions or additional hardware. Implemented on a commercial mm-wave radar platform, the approach significantly expanded the system’s field of view. The talk concluded with potential applications in autonomous navigation, disaster management, and joint communication and sensing systems.

Robust and efficient frontier pipelines for complex knowledge intensive tasks in the era of LLMs

Dr. Venktesh Viswanathan's talk focused on Complex Knowledge Intensive Tasks (CKIT), such as complex fact-checking and question answering, which require multistep reasoning and evidence retrieval from hybrid sources. The talk highlighted the challenges in current frontier models and large language models (LLMs) for CKIT, addressing retrieval and reasoning gaps. Dr. Viswanathan introduced efficient exemplar selection methods and neighborhood-aware retrieval approaches to enhance reasoning capabilities and solve retrieval issues in systems like RAG. The session concluded with discussions on the promising results, limitations, and future directions for advancing robust CKIT pipelines.

Traffic Peering Games in Internet Exchange Points

Prof. Koushik Kar’s talk presented a theoretical study on selfish peering by Internet Service Providers (ISPs) at Internet Exchange Points (IXPs), essential for global connectivity. The analysis focused on traffic pricing, traffic flows, and pairwise equilibrium. Prof. Kar discussed the efficiency of congestion-proportional pricing and extended the model to include port capacity purchases and queuing delays. The results showed that the worst-case efficiency at equilibrium, measured as the Price of Anarchy (PoA), is near the optimum. The talk also covered automating peering decisions using machine learning on public routing datasets and explored applications of game-theoretic methods in other domains, such as power and heat flow networks.

Redefining Caching for Generative AI

In this talk, Dr. Subrata Mitra focused on reimagining caching techniques in the context of Generative AI. Addressing the limitations of exact caching in generative models, Agarwal introduced two novel approximate caching systems. The first, NIRVANA, optimizes text-to-image diffusion models by reusing intermediate noise states, reducing GPU computation and latency. This work, published in USENIX NSDI 2024, significantly enhances computational efficiency. The second, Cache-Craft, extends approximate caching to large language models (LLMs) in Retrieval-Augmented Generation (RAG). Cache-Craft reuses precomputed key-value pairs, minimizing redundant computations and improving time-to-first-token and system throughput. This work will be published in ACM SIGMOD 2025.