ANC Seminar - Daniel Goodman
Tuesday, 31st January 2023
Why spike? Nonlinear mechanisms in multimodal perception
Abstract: Neurons in multimodal sensory areas display a variety of nonlinear responses, but the classical view of multimodal perception shows that optimal performance can be achieved with a linear system. So what are these neurons for? We show that in a simple task where you have to quickly respond to rare events (inspired by predator/prey decision-making), nonlinear integration becomes critical. We demonstrate this at several connected levels, using a Bayesian probability model, artificial neural networks, spiking neural networks trained with the machine learning-inspired surrogate gradient descent method, and in human psychophysics. At each of the model levels we find that nonlinear integration is essential, and the spiking nonlinearity is sufficient. We also find that humans and models are able to do an extreme version of this task in which a linear method only performs at chance level, and show qualitatively similar performance and reaction times. Please be warned that this is very much work in progress, so during this talk I will almost certainly say some stuff that is wrong (but it's more fun like that)!
Bio: Dan Goodman is a computational neuroscientist at Imperial College London. He designed the "Brian" spiking neural network simulator package, and co-founded the SNUFA spiking neural network community and Neuromatch organisation. His main interests are in understanding spiking neural networks and using machine learning as a tool for understanding intelligent processes (including the brain).
You can find him at:
Event type: Seminar
Date: Tuesday, 31st January 2023
Speaker(s): Daniel Goodman
Chair/Host: Matthias Hennig