| PhD Seminar

Name of the Speaker: Mr. Shubham Paul (EE19D407)
Guide: Dr. David Koilpillpai R
Venue: ESB-244 (Seminar Hall)
Online meeting link: http://meet.google.com/bmc-ubkh-fjw
Date/Time: 16th May 2025 (Friday), 10:30 AM
Title: End to End Communications using Autoencoders

Abstract :

This presentation comprises two parts. Part 1 explores the application of Fully Connected Neural Networks (FCNNs) in developing end-to-end communication systems, independent of established classical communication models or error control coding. The research is based on the principles of information theory and machine learning and examines various cost functions, stringent power constraints, and modifications to the Encoder architecture and training methodology. We introduce a novel twin-encoder architecture along with a model inspired by random coding. Part 2 of the presentation addresses the impact of interference in cellular networks, particularly for users at the edges of cells who face significant interference from neighbouring cells, modelled as a two-user interference channel. We present two innovative methodologies, TwinNet and SiameseNet, utilising autoencoders designed for encoders and decoders in interference-limited environments. The findings demonstrate that these models effectively leverage the interference structure to outperform traditional methods reliant on complete orthogonality. While coordinated transmissions and independent detection can enhance capacity, the specific benefits of data-driven models have not been thoroughly quantified. An analysis of the characteristics of the generated codewords provides insight into how these models achieve superior performance.