Professor Kim Hong-guk of GIST Awarded 'Best Paper Award' at Signal Processing Conference

Application of Fuzzy Theory to VQ-VAE

Professor Kim Hong-guk and student Park Dong-geon

[Asia Economy Reporter Junho Hwang] A research team led by Professor Hongkuk Kim of the Department of Electrical, Electronics and Computer Engineering at Gwangju Institute of Science and Technology won the Best Paper Award at the Joint Conference on Signal Processing, the top academic event in the signal processing field in Korea, on the 21st.

This year marks the 30th anniversary of the event, which serves as a platform to present and discuss various research results in the field of signal processing. The conference was jointly organized by four societies: The Korean Institute of Electrical Engineers, The Acoustical Society of Korea, The Korean Institute of Communications and Information Sciences, and The Korean Institute of Broadcast and Media Engineers, and was held online from the 24th to 25th of last month.

This conference featured a total of 76 papers presented through oral presentations, invited talks, and poster sessions, focusing on research results and theories in signal processing, including artificial intelligence technologies that are critically applied in industrial, engineering, and scientific fields.

Professor Kim presented the results of integrating fuzzy theory into VQ-VAE through his paper titled "Fuzzy Discrete Learning-Based Variational Autoencoder." VQ-VAE generates discrete latent variables, making interpretation easier and handling data distributions more effectively, which leads to better estimation of prior probabilities, and has recently been actively researched.

The research team proposed a model that passes distance information to the codebook as a linear combination to the decoder, maximizing codebook usage while creating meaningful data representation methods.

Professor Kim stated, "In-depth discussion and research on foundational deep learning technologies are more important than ever," adding, "It is significant that this method can create more meaningful data representations than existing technologies and has a very high potential to be used as transfer learning in downstream tasks."

Donggeon Park, the first author of the paper, said, "Applying fuzzy theory to VQ-VAE, which has been actively researched recently, aligned well with the interests of the Joint Conference on Signal Processing, which I believe led to winning this prestigious award," and added, "I hope to further develop related research to prove its validity and applicability and become a researcher who can be of some help to other researchers."