Low-Dimensional Dynamic Encoding in Prefrontal Cortex during Decision-Making
A growing body of evidence indicates that the inherent dimensionality of neural population activity is often much smaller than the number of neurons that can be recorded with current technology. However, it is also clear that single neurons often represent multiple kinds of information simultaneously, a phenomenon known as "mixed selectivity", indicating that neural population activity represents complex mixtures of sensory, cognitive, and behavioral variables. In this talk, I will describe new statistical techniques for uncovering low-dimensional latent structure from high-dimensional neural datasets. Our work represents an extension of "Targeted Dimensionality Reduction" (Mante et al, 2013), which seeks to identify subspaces that carry information about distinct task variables. We have applied our method to neural population data from macaque prefrontal cortex during a context-dependent perceptual discrimination task. It reveals the existence of independent multi-dimensional subspaces of neural activity space devoted to the coding of sensory, context, and decision-related variables on multiple timescales. I will discuss implications of this and related approaches for understanding the dimensions of neural activity underlying complex behaviors.