CNI CONNECT

Monthly Newsletter from CNI

Issue - April 2025

Upcoming Events

What’s cooking at CNI?

CNI PhD Scholar Indrasish Chakraborty’s Research Accepted at IEEE ANTS 2024

CNI PhD Scholar, Indrasish Chakraborty, has developed an innovative communication framework that integrates Simultaneously Transmitting and Reflecting (STAR) Intelligent Reflecting Surfaces (IRS) with Orthogonal Time Frequency Space (OTFS) modulation—two promising technologies poised to play a pivotal role in 6G wireless systems.

This work is the first known attempt to model and optimize STAR-IRS-aided OTFS systems for high-Doppler environments. The proposed approach includes two algorithms that jointly optimize the amplitude and phase of STAR-IRS elements, enabling efficient communication with receivers on both sides of the IRS. Simulation results demonstrate that this system significantly outperforms conventional IRS-based OTFS designs, even when constrained by practical hardware limitations such as 3-bit phase shifts.

Further details can be found in the full paper, which was presented at the 2024 IEEE International Conference on Advanced Networks and Telecommunications Systems (ANTS).

Enabling and Exploiting High-Speed In-Network Computing

Dr. Vishal Shrivastav delivered a compelling talk on overcoming hardware limitations in programmable network devices for in-network computing. He introduced Seer, a caching system that uses network delays to boost memory efficiency, and Leo, a lightweight ML-based framework for real-time traffic analysis at terabit speeds. The talk offered sharp insights into building practical, high-performance systems on constrained network hardware.

Efficient LLM Inference with HiRE and Tandem Transformers

Dr. Praneeth Netrapalli presented innovative techniques to accelerate LLM inference by addressing memory and sequential bottlenecks. He introduced HiRE, which uses dimensionality reduction and quantization to focus computation on key elements—significantly improving softmax and feedforward layer efficiency without sacrificing accuracy. He also described Tandem Transformers, a hybrid model that pairs a small autoregressive model with a large block-mode model to achieve faster and more accurate predictions. The talk offered practical insights into scaling LLMs for real-world efficiency.

Model-Assisted Optimal Control for Split Air Conditioners (ACs)

Prof. Vinayak Naik discussed energy optimization for ductless split AC systems using a model-assisted optimal control (MaOC) algorithm. The approach leverages thermal modeling and system efficiency to generate optimal cooling schedules, balancing energy consumption and cooling speed. In real-world tests, MaOC outperformed greedy and RL-based strategies, reducing energy use by 17% and cooling time by 34%. The talk also touched upon anomaly detection for predictive maintenance. This work, done with PhD student Keshav Kaushik, won the 2024 Best CS Dissertation Award at BITS Pilani, Goa.

Can kernel machines be a viable alternative to deep neural networks?

Prof. Parthe Pandit explored whether kernel machines can offer a principled and scalable alternative to deep neural networks. Motivated by the Neural Tangent Kernel and its connection to wide neural nets, he presented two key advances: data-dependent supervised kernels that adapt to the task at hand, and fast, scalable training algorithms for large-scale settings. The talk offered a refreshing take on classical ML models with modern capabilities, backed by recent theoretical and empirical results.

Decision Referrals in Human-Automation Teams: Impact of task load on team performance

Dr. Kesav Kaza, Research Fellow at the University of Ottawa, delivered a fascinating talk on the impact of task load on human-AI collaboration. He introduced a framework for optimal decision referrals in human-automation teams performing binary classification tasks, where AI handles the majority of tasks but refers some to human operators based on task load. Dr. Kaza presented a ranking scheme and algorithm for selecting tasks to refer, showing how this approach can reduce human cognitive load and improve team performance. The talk also explored applications in counter-drone surveillance and discussed open challenges in the field.