Two weeks ago, I visited the Bernstein conference in Berlin. I had lots of fun, particularly at the poster sessions, where I met William, Movitz, and Shervin. I met with each of them later and recorded the following conversations (on bark benches again^^).
William Walker (Gatsby Computational Neuroscience Unit, London) had a poster on 'Representations of State in Hippocampus Derive from a Principle of Conditional Independence'. We discuss how current deep learning struggles with generalization, lacks priors, and could benefit by learning latent conditionally independent representations (similar to place cells).
Movitz Lenninger (KTH Royal Institute of Technology, Stockholm) had a poster on 'Minimal decoding times for various shapes of tuning curves'. He was puzzled why neurons with periodic tuning curves (such as grid cells) are so rare in the brain considering their superior accuracy. He posits there may be a trade-off between accuracy and encoding time.
Shervin Safavi (Max Planck Institute for Biological Cybernetics, Tübingen) had a poster on linking efficient coding and criticality. We introduce those concepts and talk about why noise is a feature, not a bug. Shervin is also starting a new lab at TU Dresden, where he wants to understand the computational machinery of cognitive processes and he is looking for interdisciplinary-minded applicants! For Apple Podcast users, find books/papers links at: https://akseliilmanen.wixsite.com/home/post/pod05
Not familiar with place, grid and head direction cells? Here is my 5min primer.
(00:00:00) - Intro
(00:02:53) - William Walker
(00:32:53) - Movitz Lenninger
(00:55:04) - Shervin Safavi