CNI CONNECT

Monthly Newsletter from CNI

Issue - June 2025

We are pleased to announce an upcoming talk by Prof. Vikram Krishnamurthy from Cornell University on the topic of "Social Learning and Inverse Reinforcement Learning". The talk will be held on 10th June 2025. Stay tuned for further details on the venue and timings.

Registrations for the CNI Summer School 2025 on "Rumours, Consensus, and Epidemics on Networks", led by Prof. Ayalvadi Ganesh, will close on June 7, 2025. If you haven’t registered yet, this is your final opportunity to be part of this exciting program. Click here to secure your spot!

CNI Hosts Prof. Alex and Pierre for Engaging Research Interactions

In May 2025, Prof. Alexandre Reiffers Masson, an associate professor from IMT Atlantique and his student Pierre visited IISc Bangalore and engaged with the research community through insightful discussions on their ongoing projects.

As part of their visit, Alex delivered a tutorial on Reinforcement Learning algorithms, covering advanced topics such as restless bandits, weakly coupled Markov Decision Processes (MDPs), and policy design approaches including linear programming (LP)-based methods.

The visit provided valuable perspectives on current challenges and methodologies in sequential decision-making and learning.

New trends in high-dynamic range analog-to-digital converters 

In this talk, Prof. Satish Mulleti from IIT Bombay, presented a tutorial on modulo-folding ADCs, a promising approach for digitizing high dynamic range (HDR) signals using low-DR hardware. He explained how a modulo operation prevents clipping by folding the input into a smaller range, followed by digital unfolding to recover the original signal. The talk covered the mathematical foundations, unfolding algorithms, and connections to classical sampling techniques like delta and sigma-delta modulation. Prof. Mulleti also discussed hardware design and power trade-offs, concluding with a case study in digital communications to illustrate the method’s practical advantages.

Blueprint: A Toolchain for Highly-Reconfigurable Microservices

In this talk, Vaastav Anand, a PhD student, from MPI-SWS, introduced Blueprint, a toolchain designed to streamline the development and maintenance of microservice applications. He explained how Blueprint enables rapid Configure-Build-Deploy (CBD) cycles, allowing users to easily reconfigure designs, reproduce issues, and prototype solutions with minimal code changes. Supporting diverse design dimensions and major benchmarks, Blueprint simplifies key microservice workflows and accelerates iteration through an open-source, extensible framework.

Optimal Algorithms for Online Convex Optimization with Adversarial Constraints

In this talk, Prof. Abhishek Sinha from TIFR, Mumbai addressed a long-standing open problem in constrained online convex optimization (COCO). He presented the first online policy that achieves both O(\sqrt{T}) regret and O(\sqrt{T}) cumulative constraint violation, without restrictive assumptions. For strongly convex cost functions, the regret improves to O(log⁡T). His approach combines AdaGrad's adaptive regret bounds with Lyapunov optimization, leading to a surprisingly simple and elegant analysis.

Gaussian certified unlearning in high dimensions

In this talk, Aaradhya Pandey a doctoral candidate from Princeton University, explored the challenge of machine unlearning in high-dimensional settings. He analyzed a Newton-based unlearning algorithm and introduced the concept of Gaussian certifiability, showing that a single Newton step with calibrated Gaussian noise suffices to ensure both privacy and accuracy. This result contrasts prior work that required at least two steps, highlighting how the choice of certifiability definition affects theoretical guarantees in high dimensions.