Title: Communication Algorithms via Deep Learning


Speaker: Hyeji Kim (Samsung AI Research)


The design of codes for communicating reliably over a statistically well-defined channel is an important endeavor involving deep mathematical research and wide-ranging practical applications. In this talk, we demonstrate that the discovery of decoding and coding algorithms can be automated via deep learning. We first show that creatively designed and trained Recurrent Neural Network (RNN) architectures can decode well known sequential codes such as convolutional and turbo codes with close to optimal performance on the additive white Gaussian noise (AWGN) channel, which itself is achieved by the Viterbi and BCJR algorithms. We also demonstrate robustness and adaptivity to deviations from the AWGN setting. Next, we present the first family of codes obtained via deep learning which significantly outperforms state-of-the-art codes. By integrating information theoretic insights into our design of recurrent-neural-network based encoders and decoders, we are able to construct the first set of practical codes for the Gaussian noise channel with feedback. Up until now, feedback has been known to theoretically improve the reliability of communication, but no practical codes have been able to do so.


Hyeji Kim is a researcher at Samsung AI Research Cambridge in the United Kingdom. Before she joined Samsung AI Research, she worked as a postdoctoral research associate at University of Illinois at Urbana-Champaign. She received her Ph.D. and M.S. degrees in Electrical Engineering from Stanford University in 2016 and 2013, respectively, and her B.S. degree with honors in Electrical Engineering from KAIST in 2011. Her research interests include information theory, machine learning, and the interplay between the two areas. She is a recipient of Stanford Graduate Fellowship and participant of the Rising Stars in EECS Workshop in 2015.