Abstract: This talk presents our attempts at using machine learning for agricultural applications. The focus is on using vision foundation models for segmenting aerial multispectral images and crop pest classification using RGB images. We address problems related to limited labelled samples or noisy labels, which are typical of agricultural images. We also present our attempt at building a framework for monitoring, prediction and providing advisory for farmers using machine learning. Part of this work is in collaboration with Technology Innovation Hub (TIH) at IIT Bombay, for developing a mobile application that can be used by farmers.

Bio: Rajbabu Velmurugan is a professor in the Department of Electrical Engineering, Indian Institute of Technology Bombay. He received his B.E. from Govt. College of Tech. Coimbatore in 1995, M.S from Clarkson University, USA in 1998, and Ph.D. from Georgia Inst. of Tech., USA in 2007. Rajbabu's research interests are broadly in signal processing, including image, video, speech, and audio processing. His recent research focus has been on machine learning and vision for various domains (including agriculture, industrial manufacturing) and joint audio-visual processing. .  

Abstract: We consider the problem of reliable data transmission within a finite time horizon T, over a discrete memoryless channel with an unknown transition kernel. Focusing mainly on a Binary Erasure Channel (BEC) with unknown erasure probability δ, we consider a feedback model wherein the transmitter can query the receiver infrequently to obtain the empirical erasure rate experienced by the latter. A learning vs. exploitation dilemma manifests in this scenario—specifically, the transmitter needs to balance between (i) learning the channel with better accuracy and (ii) utilizing the channel to transmit as many information bits as possible. We aim to minimize a regret quantity, i.e. how much worse a learning-transmission strategy performs compared to a "Polyanskiy-Verdú-Poor" oracle who knows the channel, while operating at the same average block error rate (BLER). We propose two strategies: (i) a two-phase approach using channel estimation followed by transmission that achieves an O(T2/3) regret using only one query, and (ii) a windowing strategy using geometrically increasing window sizes that achieves an O(√T) regret using O(log(T)) queries. We further show using Le Cam's inequality that O(√T) is a lower bound on the regret attainable by any policy that depends on learning the channel—implying that (i) the geometric windowing strategy is order optimal, and (ii) the "loss in rate due to learning" is of the same order as the oracles' own inevitable loss due to channel dispersion. We conclude with remarks on feasible extensions to other channel models, including the BSC.

Collaborative work with Haricharan Balasundaram. Based on 2507.08599v2.pdf (accepted to Information Theory Workshop 2026).

Bio: Krishna Jagannathan obtained his B. Tech. in Electrical Engineering from IIT Madras in 2004, and the S.M. and Ph.D. degrees in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology (MIT) in 2006 and 2010 respectively. During 2010-2011, he was a visiting post-doctoral scholar in Computing and Mathematical Sciences at Caltech, and an off-campus post-doctoral fellow at MIT. Since November 2011, he has been with the Department of Electrical Engineering, IIT Madras, where he is currently a professor. His research interests lie in the stochastic modeling and analysis of communication networks, information theory, and queuing theory.  

Abstract: We address the classical problem of constructing confidence intervals (CIs) for the mean of a distribution, given N i.i.d. samples, such that the CI contains the true mean with probability at least $1- \delta$, where $\delta \in (0,1)$. We characterize three distinct learning regimes based on the minimum achievable limiting width of any CI as the sample size $N_{\delta}$ and $\delta \rightarrow 0$. In the first regime, where $N_{\delta}$ grows slower than $\log(1/\delta)$, the limiting width of any CI equals the width of the distribution's support, precluding meaningful inference. In the second regime, where $N_{\delta}$ scales as $\log(1/\delta)$, we precisely characterize the minimum limiting width, which depends on the scaling constant. In the third regime, where $N_{\delta}$ grows faster than $\log(1/\delta)$, complete learning is achievable, and the limiting width of the CI collapses to zero, converging to the true mean. We demonstrate that CIs derived from concentration inequalities based on Kullback--Leibler (KL) divergences achieve asymptotically optimal performance, attaining the minimum limiting width in both sufficient and complete learning regimes for distributions in two families: single-parameter exponential and bounded support. Additionally, these results extend to one-sided CIs, with the width notion adjusted appropriately. Finally, we generalize our findings to settings with random per-sample costs, motivated by practical applications such as stochastic simulators and cloud service selection. Instead of a fixed sample size, we consider a cost budget $C_{\delta}$, identifying analogous learning regimes and characterizing the optimal CI construction policy.  

Bio: Sandeep Juneja is a Professor of Computer Science and the founding director of the Safexpress Centre for Data, Learning and Decision Sciences at Ashoka University. He is also a Visiting Professor in the EE Department at IIT Bombay. Previously, he served as a Senior Professor and Dean at the School of Technology and Computer Science (STCS) at TIFR Mumbai. His research focuses on applied probability, including sequential learning, mathematical finance, Monte Carlo methods, and game-theoretic analysis of queues. More recently, his work includes epidemiological modeling and applying machine learning to monsoon weather modeling in India. He currently serves as the Area Editor for Operations Research in Simulation and has been on the editorial boards of Stochastic Systems, Mathematics of Operations Research, Management Science, and ACM TOMACS. 

Abstract: Classical model-based optimization algorithms offer provable convergence and high interpretability for solving inverse problems in signal processing, but they often suffer from slow convergence and require meticulous hyperparameter tuning. While "algorithm unrolling" has emerged as a powerful paradigm to bridge this gap—transforming iterative solvers into deep, layer-wise neural networks that learn from data—these architectures remain inherently rigid when faced with shifting real-world environments.

This talk explores the critical, yet understudied problem of domain adaptation in unrolled regression networks. Standard transfer learning techniques fail to handle continuous, parametric domain shifts governed by underlying physics, such as fluctuating noise levels, phase variations, or sensor calibrations.

To address this, we present a systematic framework that exploits the structural transparency of unrolled networks. By isolating a minimal subset of low-complexity, domain-sensitive tunable parameters, we enable precise model adaptation without the need for resource-intensive retraining. We introduce two lightweight mechanisms: Parametric Tunable-Domain Adaptation (P-TDA), which incorporates known environmental parameters at inference, and Data-Driven Tunable-Domain Adaptation (DD-TDA), which utilizes an auxiliary network to infer hidden domain variations directly from raw inputs. Finally, we demonstrate the efficacy of our framework across three distinct sparse recovery bottlenecks—noise-adaptive sparse reconstruction, blind gain sensor calibration, and non-linear phase retrieval—showing that our architectures match the performance of dedicated specialist networks while maintaining a drastically lower parameter footprint. 

Bio: Dr. Satish Mulleti is currently an Assistant Professor of Electrical Engineering at IIT Bombay. He earned his M.Tech. from IIT Kanpur and his Ph.D. from the Indian Institute of Science (IISc), Bangalore, followed by a four-year postdoctoral tenure in Israel at both the Technion and the Weizmann Institute of Science. He worked as a Researcher at ISRO and TCS Innovation Labs. His research focuses on bridging classical signal processing with modern data science, specializing in deep learning theory, compressed sensing, finite-rate-of-innovation sampling, blind deconvolution, and spectral estimation. 

Abstract: With significant increase in live video traffic in cellular networks, multicast/broadcast adaptation is being considered seriously. Standardization bodies like 3GPP and ATSC are looking to improve the multicast/broadcast services. In this talk, we will talk about some resource allocation problems in such integrated networks and point to some open problems in this area.

Bio: Prasanna Chaporkar is a Professor in the Department of Electrical Engineering at Indian Institute of Technology Bombay and heads the Information Networks (Infonet) Lab. His research focuses on next-generation communication networks, including wireless and mobile networks, 5G/6G systems, network optimization, stochastic control, resource allocation, distributed algorithms, and sustainable networking. 

He obtained his B.Tech. from Walchand College of Engineering, M.Sc. (Engineering) from Indian Institute of Science, and Ph.D. from University of Pennsylvania. He joined IIT Bombay in 2007 and has made significant contributions to wireless communications, networking, and optimization research. 

Prof. Chaporkar has authored over 100 research publications, holds several patents, and has contributed to national and international wireless communication standards, including work on beyond-5G and 6G technologies. His research group actively collaborates with industry and government organizations on cutting-edge networking technologies.

Abstract: Many optimization problems today are solved not by one computer but by a network of agents, each possessing only part of the data and communicating only with its neighbors. How fast can such a network optimize? And what limits that speed?

This talk explores distributed optimization through two complementary viewpoints. First, we show how continuous-time primal–dual dynamics can exhibit accelerated convergence through an exact energy conservation law, and why this acceleration may disappear after seemingly natural discretization. A lower-bound result reveals a fundamental limitation of a broad family of single-loop schemes and motivates algorithms that deliberately trade additional communication for acceleration. Second, we develop an accelerated gradient-tracking viewpoint in which the average network dynamics become an inexact version of Nesterov acceleration. Chebyshev filtering suppresses the resulting network error, leading to near-optimal convergence guarantees and robustness under suitable changes in network topology.

The broader theme is that in networked optimization, computation, communication, acceleration, and robustness cannot be designed independently. The talk will develop these ideas intuitively, with examples and geometric interpretations, before presenting the main theoretical insights.

Bio:  Mayank Baranwal is a Senior Scientist with the Tata Consultancy Services (TCS) research division in Mumbai. He also holds an Adjunct appointment with the Systems and Control group at the Indian Institute of Technology, Bombay (IITB). Before joining TCS, he was a postdoctoral scholar in the Department of Electrical and Computer Engineering at the University of Michigan, Ann Arbor. He received his Bachelor in Mechanical Engineering in 2011 from the Indian Institute of Technology, Kanpur (IITK), an MS in Mechanical Science and Engineering in 2014, an MS in Mathematics in 2015, and PhD in Mechanical Science and Engineering in 2018, all from the University of Illinois at Urbana-Champaign (UIUC). His research interests are modeling, optimization, control, and inference in network systems with applications to distributed optimization, supply-chain networks, power networks, control of microgrids, bioinformatics, computational biology, and deep learning theory. Mayank is a recipient of the Institute Silver Medal in 2011 (IIT Kanpur), the ME Outstanding Publication Award in 2017 (the University of Illinois), the Young Scientist Award in 2022 (Tata Consultancy Services), and the AI Research Award in 2024 (Nasscom AI). He is also a Young Associate with the Indian National Academy of Engineering (INAE). 

Abstract: Conventional active phased arrays rely on multiple power-hungry radio frequency (RF) chains, high-speed data converters, and complex analog phase shifters, making them very expensive. This talk will introduce Space-Time-Coding (STC) digital metasurfaces as a low-cost, highly scalable alternative. We will explore how applying temporal modulation directly at the carrier level allows for the dynamic manipulation of harmonic beams. Next, the talk will discuss a fundamental paradigm shift in transceiver design. It will demonstrate how STC metasurfaces can operate entirely in the wave domain to simultaneously embed data and break physical reciprocity for highly secure communications. 

Bio: Dr. Debidas Kundu is an Assistant Professor in the Department of Electrical Engineering at Indian Institute of Technology (IIT) Bombay. He received his Ph.D. degree from the Dept. of Electronics and Electrical Communication Engineering at the Indian Institute of Technology Kharagpur, India, in June 2018. He received the DST-INSPIRE Faculty Award and subsequently joined the ECE Department of IIT Roorkee as an INSPIRE Faculty member in January 2019. From April 2022 to April 2023, he worked as a visiting researcher (Post-Doc) at Carleton University, Canada. Prior to joining IIT Bombay, he was an Assistant Professor at Indian Institute of Technology Delhi. He is a recipient of the Young Scientist Award from URSI-RCRS and URSI-GASS in 2020 and 2023, respectively and DRDO Dare to Dream Award in 2026. His current research mainly includes analytical techniques in electromagnetics, metasurfaces, reconfigurable intelligent surfaces (RIS), electromagnetic wave scattering and polarization control, and microwave engineering. 

Abstract: Indian Rummy is a widely played imperfect-information card game with a game tree estimated to contain 10^{170} decision nodes. Unlike other card games such as Poker, Bridge, and Dou Dizhu, Indian Rummy is strongly shaped by meld structures---sets and sequences---and by continuously evolving hand states produced by sequential "pick" and "discard" actions. These properties make it difficult to directly apply standard abstraction and equilibrium-computation methods to Indian Rummy. The recurring decisions in two-player Indian Rummy are whether to play or drop, whether to pick from the open or closed pile, whether to declare, and which card to discard. Beyond individual games, tournament play also introduces long-horizon complexity because decision making must account for the current score and the number of games remaining.

We propose RUDRA (a Rummy Utility-DRiven Agent), which addresses the intractability of Indian Rummy by decomposing it into subproblems. RUDRA integrates four main elements.  First, we develop a high-performance C++ environment for simulating games, which includes a fast engine to compute optimal card groupings. Second, building on this engine, we introduce RUDRA-PS, a mixture-of-experts policy trained via policy search. Third, we introduce RUDRA-SIM, a simulation-based planner that combines Monte Carlo simulation with Bayesian opponent hand prediction and RUDRA-PS to evaluate candidate actions over plausible hidden states. Fourth, we introduce RUDRA-EID to compute the initial play/drop strategy in tournaments by solving a finite-horizon imperfect-information extensive-form game. Using learned hand-strength abstractions, the game is reduced from 10^{23} states to approximately 10^{8} decision nodes and solved by combining dynamic programming and linear programming.

RUDRA combines RUDRA-SIM for pick and discard actions with an RUDRA-EID for equilibrium initial drop strategy in the tournament. We validate it by playing against state-of-the-art baselines and also through controlled human evaluations on a custom-built online gameplay platform. In five separate 320-game tournaments against experienced Indian Rummy players, RUDRA won all five tournaments, achieving an average advantage of $1.96$ points per game. These results establish RUDRA as a state-of-the-art agent for Indian Rummy. We expect both the general approach of RUDRA (to decompose a massive imperfect-information game into strategically meaningful subproblems) and its specific modules (for optimisation, learning, inference, simulation, and abstraction) to benefit agent-development for large imperfect-information games.

This work is based on the Ph.D. thesis of Santhosh Kumar Guguloth. 

Bio: Shivaram Kalyanakrishnan is an Associate Professor in the Department of Computer Science and Engineering at the Indian Institute of Technology Bombay. His research interests include artificial intelligence and machine learning, spanning topics such as sequential decision making, multi-agent learning, multi-armed bandits, and humanoid robotics. Kalyanakrishnan received a Ph.D. in computer science from the University of Texas at Austin. Subsequently, he was a Research Scientist at Yahoo Labs Bangalore and an INSPIRE Faculty Fellow at the Indian Institute of Science, Bengaluru. Kalyanakrishnan works on both theoretical and applied problems. He heads e-Yantra, a large-scale outreach programme that annually trains thousands of students in robotics

Abstract: Data centre networks require new transport protocols in view of the large bandwidth and low RTT in the network together with strict latency requirement for distributed compute. Low latency can be achieved by stall free scheduling between a sender receiver pair to ensure minimal buffering at the intermediate switches. The main technical challenge in such protocols is to establish large number of matched sender receiver pairs for concurrent message transfer. We propose a novel distributed bipartite matching algorithm with local information that improves the mean matching size in random bipartite graphs that model the data centre network communication.

Bio: Parimal Parag is currently an associate professor in department of electrical communication engineering at Indian Institute of Science at Bangalore. He was working as senior systems engineer in R&D at ASSIA Inc. from October 2011 to November 2014. He received his B. Tech. and M. Tech. degrees fromIndian Institute of Technology Madras in fall 2004; and the PhD degree from Texas A&M University in fall 2011. He was at Stanford University and Los Alamos National Laboratory, in autumn of 2010 and summer of 2007, respectively.

His research interests are in design, performance evaluation, and control of large distributed andnetworked intelligent systems applying mathematical tools from queueing theory, information theory, coding theory, and optimization methods.