Synapse

 

Monthly Newsletter from CNI

Issue - August 2023

From the Editor

“Welcome to the August edition of Synapse - the monthly newsletter from Centre for Networked Intelligence.

This month, IISc welcomes new students, with classes commencing from the first week of August. We have weekly seminars from CNI this month from various researchers across the globe. CNI will be welcoming new scholars to the centre and yet another year of groundbreaking research activities starts here.”

PAC Mode Estimation using PPR Martingale Confidence Sequences by Shivaram Kalyanakrishnan from IIT Bombay on August 2nd @ 4pm

Almost-optimal best restless markov arm identification with fixed confidence from Dr. P N Karthik from NUS, Singapore on August 17th @ 5pm

Vincent Y. F. Tan, a National University of Singapore associate professor, presented the " Probably Anytime-Safe Stochastic Combinatorial Semi-Bandits " problem. Agents choose item subsets, aiming to balance rewards and risks over time. Tan introduced PASCOMBUCB, an algorithm minimizing regret and risk in decision-making. Applicable to recommendation systems and transportation.

Microsoft researcher Greg Yang discussed the vital role of mathematics in improving large-scale deep learning's effectiveness. He addressed stability, proposed transferring hyper-parameters, and emphasised numerical stability, normalisation, and gradient concepts. Yang highlighted math's power in optimising and enhancing deep learning models, urging a comprehensive theory for understanding their workings.

Aurélien Garivier from Ecole Normale Supérieure de Lyon addressed missing mass and optimal discovery , focusing on probability estimation in scenarios with limited data. He presented statistical methods like the Jackknife estimate for calculating unseen event probabilities and explored the balance between exploration and exploitation in optimal discovery strategies. The talk encompassed Bandit algorithms, time considerations, and stressed the significance of feedback and knowledge in the discovery process.

Siva Theja Maguluri, a professor at Georgia Tech, discussed tail probability bounds for queue lengths in Heavy-Traffic queueing systems. Exponentially decaying bounds were provided for P(ϵq > x), where ϵ represents proximity to max load. Applicable for finite ϵ, getting sharper as ϵ → 0, using exponential Lyapunov function and Markov’s inequality. Demonstrated approach with examples, offering improved tail bounds for queue lengths in such systems.

Sarath AY, a postdoctoral researcher at Brown University, established a Sanov-type large deviation principle for empirical measures of marked unimodular random graphs like Erdos-Renyi and random regular graphs. The rate function involves relative entropy, leading to a tractable form. The work also yielded a variational formula for annealed pressure in statistical physics models on sparse random graphs.

Sanjay Shakkottai from the University of Texas at Austin addressed non-convex optimisation in machine learning for realistic functions . His technique ensures accurate solutions despite non-uniformly bounded data noise, crucial for practical scenarios. The algorithm's effectiveness was demonstrated for various functions, outperforming certain previous approaches for specific function types.

Sushil Varma, a Ph.D. student at Georgia Tech, discussed dynamic pricing and matching in online marketplaces . He addressed challenges in aligning supply and demand, focusing on profit maximization and customer service speed. The talk covered centralized pricing, optimization strategies, stability, and compatibility. Emphasis was placed on considering internet structure, balancing profit and waiting time, and challenges like system state distribution and server incorporation.

Dr. Pannag R Sanketi, a Tech-Lead Manager in the Robotics team at Google DeepMind presented an overview of the team’s robotics research. They aim to create a versatile robot model, scaling size, power, and data for complex tasks. Open-source datasets for model fine-tuning, discussing labelling, simulation, reinforcement learning, and projects like navigation, agility, and language-based planning. Also, integrating manipulation data into large models for independent task execution.

Mahima Agumbe Suresh, an Assistant Professor at San Jose State University, presented SaFeR, a safety framework for e-scooter riders using Advanced Driver Assistance algorithms for real-time obstacle awareness. Depth perception via monocular cameras and optimisation techniques were explored. SaFeR reduces network complexity, emphasises integration, collaboration, and safety enhancement, addressing challenges like early evasive action and regulatory compliance in e-scooters.

Suryanarayana Sankagiri, a postdoctoral researcher at EPFL, explored blockchain robustness to network delays, discussing security guarantees, challenges in analysis, and proposing a model with random delays. He emphasised confirmation processes, vulnerabilities, and mathematical assumptions, suggesting relaxed network assumptions for similar results. Sankagiri offered insights into blockchain security analysis and network conditions.