A Framework for Neuroevolution
Taliesin Benyon (Wolfram Research, University of Witswatersrand)
Hosted by NERV
Artificial neural networks (ANNs) trained with deep learning can perform complex tasks, similar to those performed by organisms with nervous systems. Curiously, in the biological world, some behaviours are not learned but genetically hard-coded: mice respond to looming stimuli, hatched turtles instinctively head out to sea, and babies in the womb appear to prefer face-like visual stimuli. Classic deep learning does not model these phenomena well. This talk will discuss a framework to tackle the emergence of such innate behaviours, as well as more general phenomena of neural systems.