A1867
Title: Data-driven and normative approaches to modeling neural data
Authors: Jacob Yates - University of California, Berkeley (United States) [presenting]
Abstract: Modern computational neuroscience has developed two complementary ways of building statistical models of neural activity. Data-driven approaches start from neural recordings and seek flexible descriptions of the relationship between stimuli, behavior, latent state, and spiking responses. Their strength is empirical fidelity: they can capture structure in large neural populations without requiring that the model itself solve the animal's task. Normative approaches begin from the opposite direction. They train artificial systems to solve perceptual or inferential problems, then ask whether the resulting representations, dynamics, or errors resemble those observed in the brain. Recent work on Poisson variational autoencoders offers a bridge between these approaches. Poisson models are a natural statistical language for spiking data, but they can also define latent generative models that perform inference. This shared statistical structure makes it possible to build models that are both useful tools for describing neural responses and candidate accounts of the computations those responses implement.
