Decoding Visual Stimuli from Cortical Activity using Neural Networks

Abstract

This thesis explores the application of deep learning techniques for reconstructing visual stimuli from neural activity in the primary visual cortex (V1). The focus is placed on overcoming the scarcity of biological data by developing data-efficient architectures, analyzing the impact of synthetic training data, employing adversarial and transfer learning, and introducing novel auxiliary optimization objectives. A series of experiments is conducted using data from in silico simulations of cat V1 and in vivo recordings from mouse V1, highlighting the best-performing decoding approach and offering suggestions for future research. Notably, the methods developed in this thesis outperform some existing state-of-the-art decoding techniques according to several widely used evaluation measures. Overall, the results underscore the potential of machine learning in neural activity decoding and pave the way for future advancements in brain-computer interfaces and neuroscientific research.

Jan Sobotka

CS Master’s Student & AI/ML Research Assistant

I am a master’s student in computer science at EPFL, and a research assistant at the MLBio Lab. I am interested in representation learning, (mechanistic) interpretability, meta-learning, reasoning, test-time training, and machine consciousness.