CNI CONNECT

Monthly Newsletter from CNI

Issue - September 2025

Join us for the today’s talk on “Decoupling theorem in quantum information theory: Application and Derandomization” by  Prof. Aditya Nema, (IIT Gandhinagar) on 22 September 2025.

Join us for the upcoming talk on “Distributed Caching of Transient Contents in a Tandem of Caches” by Ms. Ankita Koley (Indian Institute of Science) on 30 September 2025.

The Centre for Networked Intelligence (CNI), in collaboration with ECE and RBCCPS at IISc, hosted a workshop on Quantum Networking by Prof. Thirupathaiah Vasantam from Durham University, UK.

Prof. Vasantam introduced the foundations of quantum networking—covering concepts such as qubits, superposition, entanglement, and quantum gates—and highlighted milestones including quantum key distribution, Shor’s algorithm, and satellite-based experiments. He discussed the challenges of building scalable quantum networks, including photon loss, error correction, and the no-cloning theorem, and explained how techniques like quantum teleportation and entanglement swapping pave the way for long-distance communication.

The talk also emphasized the potential of distributed quantum computation, where multiple smaller quantum processors are interconnected to form large-scale “quantum data centers.” Prof. Vasantam encouraged greater participation from the engineering and computing communities to address the open challenges and advance this emerging interdisciplinary field.

We are pleased to announce that CNI scholar Ankita Koley’s paper, “Distributed Caching of Transient Contents in a Tandem of Caches”, has been accepted for presentation at the Allerton Conference on Communication, Control, and Computing (UIUC, Sept 17–19, 2025).

Her research tackles the challenge of caching short-lived content across multiple caches arranged in tandem, where fetching data from upstream caches or servers incurs significant costs. By framing the problem as a continuous-time Markov decision process and leveraging Whittle index-based policies, she develops a distributed caching strategy that eliminates the need for centralized coordination while maintaining performance close to optimal.

A Tutorial on Generative AI Using Diffusion Models

Prof. Sanjay Shakkotai (University of Texas at Austin) presented a tutorial on generative AI with diffusion models, explaining their mathematical foundations such as stochastic differential equations, forward–reverse processes, and score-based learning. He also highlighted recent advances enabling applications in image editing, stylization, and personalization.

Scaling AI Model Serving: QoS, Multimodality, and Beyond

Dr. Jayashree Mohan (Microsoft Research) discussed the challenges of scaling AI model serving and introduced two systems addressing them. Niyama improves QoS-driven inference for large language models through fine-grained scheduling and overload management, while ModServe enables efficient serving of multimodal models via modular optimization and autoscaling. Together, these systems significantly enhance throughput, reduce costs, and ensure latency guarantees in large-scale deployments.

Equilibrium Cycle: A "Dynamic" Equilibrium, and Its Application in Ride Hailing

Prof. Jayakrishnan Nair (IIT Bombay) introduced the concept of an equilibrium cycle, a dynamic counterpart to the Nash equilibrium that captures oscillatory game dynamics. This set-valued solution applies even to discontinuous games and generalizes classical concepts like minimal curb sets. He illustrated its properties through examples from economics and highlighted its application in modeling competition between ride-hailing platforms.

Efficient Solutions for Machine Learning at the Edge

Prof. Saurav Prakash (IIT Madras) discussed efficient and privacy-preserving machine learning at the edge. He highlighted challenges posed by resource-constrained, heterogeneous devices and limited data sharing, and presented solutions enabling federated learning of large global models using smaller local models on edge devices.

BanditSpec: Adaptive Speculative Decoding via Bandit Algorithms

Prof. Vincent Y. F. Tan (National University of Singapore) presented BanditSpec, a novel framework for adaptive speculative decoding of large language models. By framing hyperparameter selection as a multi-armed bandit problem, his approach dynamically optimizes decoding during text generation without prior training. Experiments with LLaMA3 and Qwen2 show that BanditSpec achieves near-optimal throughput, outperforming existing methods in real-world LLM serving scenarios.