Comparing computations across brains and networks
When an animal makes a decision, or its internal state changes, the activity of large neural populations evolves over time. These population dynamics commonly unfold on low-dimensional neural manifolds. To understand the computations they implement, we need methods that learn the dynamical processes on these manifolds and turn them into interpretable latent representations that are consistent across animals, recording sessions and artificial networks.
Our approach
MARBLE is an unsupervised representation learning method for neural dynamics. It decomposes the dynamics on the neural manifold into local flow fields and maps them into a common latent space using geometric deep learning. More generally, it introduces a statistical learning paradigm for non-linear dynamical systems, based on decomposing the dynamical attractor into local flow fields.
What we found
We applied MARBLE to simulated nonlinear dynamical systems, recurrent neural networks, and single-neuron recordings from primates and rodents. In each case, it discovered emergent low-dimensional latent representations that parametrise high-dimensional neural dynamics during gain modulation, decision-making and changes in internal state. These representations are consistent across neural networks and animals, which enables robust comparison of cognitive computations.
In extensive benchmarks, MARBLE achieved state-of-the-art within- and across-animal decoding accuracy compared with current representation learning approaches, with minimal user input. Our results suggest that manifold structure is a powerful inductive bias for developing decoding algorithms and for assimilating data across experiments.
The code is open source on GitHub. For a short overview, see our Research Briefing in Nature Methods.