Speaker: Xuexin Wei ARNI WG Multi-resource-cost optimization of neural network models

Title: Constraints of efficient neural computation

Abstract: Neural systems adapt to the statistical structure of the environment to support behavior. While it is generally recognized that such adaptation is subject to various biological constraints (such as noise, metabolism, wiring cost), how these constraints determine the optimal neural computation remains unclear. For the first part of this talk, I will discuss theories of efficient coding based on consideration of metabolic cost and neural noise. For the second part, I will present ongoing work on how the geometry of the stimulus manifold shapes the structure of neural code. In particular, using the processing of heading direction as an example, I will show that the asymmetry of the stimulus manifold naturally accounts for key properties of heading direction encoding in macaque MST.