Interpretability is the new frontier in AI research. Understanding how generative models learn and how they resemble or differ from humans can bring insights for diverse fields such as neuroscience and decoding animal communication. In this talk, I present several techniques for introspecting deep neural networks. I also propose a model called ciwaGAN that features several aspects of human language acquisition that other models lack (embodiment, communicative intent, production-perception loop). Together, interpretability techniques and realistic models of human language can bring us closer to answering some of the fundamental questions about human and non-human language learning. Using the proposed techniques, we can compare and evaluate artificial and biological neural processing of language as well as discover meaningful patterns in data as unknown as that of whale communication.
Gašper Beguš is an Assistant Professor at the Department of Linguistics at UC Berkeley where he directs the Berkeley Speech and Computation Lab. He is also the Linguistics Lead at Project CETI and a Member of Berkeley's Institute of Cognitive and Brain Sciences. His research combines machine learning and statistical modeling with neuroimaging and behavioral experiments to better understand how deep neural networks learn internal representations and how humans learn to speak.