The analysis of neural population activity during behaviour consistently uncovers low-dimensional mathematical structures that capture a large fraction of neural variability. These structures or “neural manifolds” are defined by the dominant patterns of covariation across neurons. Recent studies focusing on neural manifolds and the activity within them –the “latent dynamics”– have shed light into questions about cognition, motor control, and learning that had long remained elusive.

In this talk, I will discuss our ongoing work to understand the emergence of neural manifolds and their role in the generation of behaviour using a combination of neural recordings in behaving animals, and computational models. First, I will present a recent study where we tested the hypothesis that due to genetically specified constraints on circuit architecture, different individuals from the same species that are engaged in the same behaviour would produce preserved latent dynamics.

Next, I will discuss some of our research to understand the processes implemented by these latent dynamics. I will focus on a specific timescale of behaviour: adaptation of a learned skill. I will present a novel bottom-up model of motor adaptation in recurrent circuits performing online error correction, and show that it readily recapitulates many electrophysiological and behavioural results in motor learning.

Overall, these studies support the notion that neural manifolds are “objects” with meaningful functions in the generation of behaviour, whose properties are shaped by functional and biophysical constraints on neural circuits.