PhD Defence | Biologically Plausible Deep Learning: Should Airplanes flap their Wings?
Peter O’Connor’s thesis examines how neural networks can effectively be trained while obeying biological constraints. O’Connor completed his research under the supervision of Max Welling and Efstratios Gavves (both from UvA).
Deep neural networks follow a pattern of connectivity that was loosely inspired by neurobiology. The existence of a layered architecture, with deeper neurons representing increasingly abstract features, was known from neuroscience long before it was used in machine learning. However, when one looks beyond superficial similarities, deep networks appear to be fundamentally different to their biological counterparts in the way they communicate, the way they learn, and the domain in which they operate.