Speaker
Description
Neural codes must support different kinds of computation. In particular, they need to preserve latent relations among experimental and task conditions through abstraction, so that a rule learned in one context can transfer to another, and compositionality, so that knowledge of observed feature combinations can guide responses to previously unseen combinations. At the same time, they must be flexible in order to simultaneously implement many tasks through simple downstream readouts. We develop a high-dimensional theory based on a model that preserves key aspects of latent task geometry, neural encoding and trial variability, while remaining analytically solvable in the limit of infinitely many neurons. We derive exact asymptotic formulae for observables that quantify these computational capabilities, including storage capacity and cross-condition generalization performance, and show how they are controlled by geometric properties that can be measured directly from neural population recordings.