Synapse

Monthly Newsletter from CNI

Issue - March 2024

Join us at the upcoming talk on "Between Promise and Reality: Navigating the Landscape of Indic Language LLMs Amidst Hyperbole and Hope" by Mr. Gaurav Aggrawal , founder and CEO of Ananas Labs, on 12/03/2024 at 17:00 hrs IST.

CNI, in collaboration with the eBPF Foundation, is hosting eBPF Day India, a one-day workshop scheduled for March 16, 2024. The event aims to increase awareness of eBPF, showcase its use in various fields such as networking, security, and observability, and encourage discussions on its future applications. It is open to academic/industry professionals, along with students, and will be conducted in hybrid mode.

Indian Institute of Science, Bengaluru, oraganized Annual Open Day on February 24, 2024.  All the student community and the general public had an opportunity to explore exciting science and technology initiatives, lectures, experimental demos, poster presentations, quiz contests, scientific competitions and exhibitions that were showcased in the various departments and centers across the institute. Some of the CNI members also presented their posters and demos for the Open Day.

CNI fellows, Tejashree S , Moonmoon Mohanty and Mohd. Azfar presented a poster titled “ Video Lecture Generation Using AI ” on the Open Day. It aimed at reducing the workload of teachers by automating the process of video lecture creation from textual lecture notes. It was segmented into 4 phases as follows:
1. Transcript Generation Phase: Transforms lecture notes into a summarized transcript using AI tools like ChatGPT 3.5.

2. Audio Generation Phase: Converts the transcript into speech using an AI voice generation tool like ElevenLabs, requiring input from a voice model. 

3. Video Generation Phase: Creates video content from Beamer presentations using online PDF to video converters. 

4. Sync Phase : Merges the audio with the video to produce the final video lecture, utilizing tools such as Davinci Resolve. 

The project demonstrated the integration of AI tools in educational content creation, making it possible to generate video lectures from written materials automatically.

CNI fellows, Prashanth S. and Laxmikant Pai Angle, along with their lab members, demonstrated games based on Reinforcement Learning (RL), specifically directed towards younger audiences. These games were designed to represent various degrees of complexity, determined by the number of training epochs of the RL agent, to cater to a diverse range of skill levels of participants. Students had the opportunity to challenge varying levels of the RL agents and compete for positions on the leaderboard. This interactive experience provided students with an intuitive understanding of basic RL principles and strategic decision-making, all within a competitive and educational environment.

Learning while bidding in Vickrey Auctions with acquisition rate and targeting constraints

In this talk, Prof. Ravi Mazumdar discussed how to effectively bid in auctions, using the unique setup of Vickrey auctions where you win by bidding the highest yet pay the second-highest amount. He detailed how bidders can adapt their strategies over time, considering the desired frequency of wins and the specific market segments they aim to target. By modelling it as a convex optimization problem, he demonstrated how to calculate optimal bids and allocate resources, given sufficient market supply. Stochastic approximation and dual-timeframe techniques are used to fine-tune bidding and resource allocation strategies as conditions change. He explained how to meet contract requirements even with unpredictable market supplies by continuously learning and adjusting bids and resources based on actual outcomes. At the end, through numerical simulations and theoretical results, he showcased the practical implications and effectiveness of the discussed methods in real-world scenarios.

Min LQG games and collective discrete choice problems 

In this talk, Prof. Roland P. Malhamé presented a new model for analyzing group decision-making influenced by social dynamics, with a focus on scenarios like elections and micro-robotic swarm explorations. He discussed two main scenarios: a deterministic one where individuals’ choices are fixed but influenced by peers, and a stochastic one where choices change probabilistically over time. Mathematical models confirm that in both scenarios, groups can achieve ε-Nash equilibria, indicating that decentralized decisions can lead to consistent group outcomes. Specifically, he pointed out that while deterministic settings result in static final choices, stochastic settings allow for evolving decisions that eventually reach stability as the number of participants grows. The significance of these findings lies in their application to decentralized systems like robotic swarms, illustrating that without a central command, individual components can still make unified decisions. Lastly, the model offers insights into how collective choices can swing under social pressures and how groups can reach decisions while facing communication constraints, demonstrating its broader implications to understand political dynamics, societal choices and strategic planning.

Power Allocation for Generalized Gaussian MAC

In his talk, Prof. Lakshmi Narasimhan Theagarajan, provided a comprehensive overview of the strategies for optimizing power distribution in communication systems, specifically focusing on Multiple Access Channels (MAC) with generalized Gaussian noise. After a brief overview of the system setup and challenges, he delved into mathematical formulations, highlighting key metrics like sum-rate capacity and system performance bounds. He compared the two approaches, Centralized vs Greedy power allocation strategies demonstrating the conditions under which the greedy approach is particularly effective, such as infinite transmissions and specific power levels. He also highlighted the significance of randomized power allocation to address the non-convexity of the sum-rate function, promoting system efficiency. He concluded the talk with empirical analysis, comparing the performance of different allocation algorithms (Alternating Maximization Algorithm vs. Greedy vs. Uniform) and their effectiveness as the number of users vary.

Online Age of Information Scheduling

Mr. Kumar Saurav discussed strategies to keep information fresh in networked systems, highlighting its relevance in fields like tele-robotics and virtual reality. He highlighted the complexities involved in scheduling updates to minimize the Age of Information (AoI), addressing challenges such as transmission delays, energy constraints, and coordinating multiple sources. He pointed out the shortcomings of conventional packet-based metrics, such as latency, and the insufficiency of older scheduling strategies like Shortest Remaining Processing Time (SRPT) in meeting AoI goals. Later, he proposed advanced scheduling strategies such as Greedy Policy and an improved SRPT+ to enhance size and generation time of updates. A novel randomized scheduling policy that addresses transmission costs and AoI in a multi-source setting is explored. It involves two main subroutines: update selection and source scheduling to balance transmission efficiency and information freshness.

A decentralised algorithm for minimizing multi-agent congestion cost on a network

In this talk, Prof. Nandyala Hemachandra introduced a Multi-Agent Congestion Cost Minimization (MACCM) approach, a decentralized learning algorithm aimed at reducing congestion costs in multi-agent systems by extending the classical Stochastic Shortest Path (SSP) framework with congestion-dependent costs. He discussed the transition from single agent to multi-agent models, highlighting the complexities introduced by private costs, such as fuel efficiency, and shared costs like congestion from agents selecting identical routes. He emphasized the importance of a decentralized approach, enabling each agent to operate independently without communicating their actions, which preserves privacy and scales effectively with the number of agents and network size. Experimental results were shared validating the algorithm's efficiency in reducing congestion costs and maintaining minimal regret across complex network scenarios.