Synapse

Monthly Newsletter from CNI

Issue - October 2024

Join us for the upcoming talk on "EdgeRIC: Empowering AI-based Real Time Intelligent Control, Optimization and Security in NextG Networks" by Prof. Dinesh Bharadia, University of California San Diego on 08/10/2024.

Join us for the upcoming talk on "Efficiently Serving Large Language Models" by Dr. Ashish Panwar, Senior Researcher, Microsoft Research India on 15/10/2024.

Upcoming DoT Workshop on Bharat 6G Vision: Bridging Research and Standards

Join leading experts at IISc Bangalore on September 2-3, 2024, to shape India's 6G future. The workshop will cover ICT standardization, intellectual property rights, and India's contributions to global telecom policies. Engage in discussions with key stakeholders from academia, industry, and government to drive innovation and strategic growth in 6G technology.

Low Complexity Optimal Policies for Networked Control Systems

 In this talk, Manali Dutta, a PhD student at the Indian Institute of Science, Bangalore, discussed optimal scheduling policies for wireless networked control systems (WNCS) operating over unreliable communication channels. She presented low-complexity solutions for three key challenges: optimizing transmissions when the channel is partially observed, designing risk-sensitive policies that account for higher-order cost moments, and handling half-duplex constraints at the controller. Their findings reveal threshold-based strategies where transmissions occur only when specific conditions on channel belief, estimation error, or plant state magnitude are met.

Schema Generation for Querying a Data Lakehouse

In this talk, Balaji Ganesan, a Senior Research Engineer at IBM India Research Lab, discussed the challenges of using Large Language Models (LLMs) for schema generation in Data Lakehouse applications. He highlighted how generating and evaluating schemas is more complex than SQL query generation due to database constraints and correctness verification. Their work explores techniques like schema linking, knowledge infusion, and uncertainty quantification, extending them to GraphQL query and schema generation.

Locality Sensitive Hashing in Fourier Frequency Domain for Soft Set Containment Search

In this talk, Prof. Abir De, IIT Bombay, discussed efficient retrieval methods for search applications involving passage retrieval, text entailment, and subgraph search. He introduced FOURIER HASHNET, a novel data-sensitive indexing method that transforms asymmetric hinge distance into a dominance similarity measure using Fourier transforms. This approach enables the use of traditional Locality-Sensitive Hashing (LSH) in the frequency domain, improving retrieval efficiency. Their experiments demonstrate that the proposed dominance similarity measure and trainable hash codes significantly enhance query time and retrieval quality compared to existing methods.

Flipped Huber: A new additive noise mechanism for differential privacy

 In this talk, Prof. Sheetal Kalyani from IIT Madras presented a novel noise addition mechanism for differential privacy that balances privacy protection and accuracy. The proposed mechanism samples noise from a hybrid density that combines the sharp center of a Laplace distribution with the light, sub-Gaussian tails of a Gaussian distribution. She provided theoretical analysis ensuring differential privacy guarantees and derived conditions for its applicability in both one and higher dimensions. Numerical simulations demonstrated that this method achieves a superior trade-off between privacy and accuracy compared to existing mechanisms.

Risk-Sensitive Bandits: Arm Mixtures Optimality and Regret-Efficient Algorithms

In this talk, Dr. Arpan Mukherjee, a postdoctoral researcher at Imperial College London, introduced a new framework for risk-aware sequential decision-making in stochastic multi-armed bandits. He explored distortion riskmetrics, a broad class of risk measures, and highlighted the novel insight that optimal strategies often involve selecting a mixture of arms rather than a single best arm. This challenges conventional bandit algorithms, which are not designed for such mixtures. To address this, he presented new algorithms capable of tracking optimal mixtures and discussed the technical challenges in establishing regret lower bounds under this setting. The talk concluded with open questions on risk-sensitive decision-making and future research directions.