Computation and Neural Manifolds

David L Barack for the Mind, Technology, and Society (MTS) Talk series on April 10, 2023.

Abstract: Recent research in cognitive neuroscience has uncovered so-called neural manifolds that play a central role in explanations of behavior. Revealed using a range of dimensionality reduction techniques, these manifolds are entities in low-dimensional spaces contained in high-dimensional neural spaces. This talk explores a possible computational interpretation for the role of manifolds in cognition. This talk argues that manifolds provide evidence for what neural computations are performed, and later argues that manifolds also provide evidence for how inputs are transformed in to outputs during neural computation. Finally, this talk explores how this computational role for manifolds constrains a conceptual analysis of computation in the brain.

[haluk]

Nice research but wrong direction. Single neurons act. Single (pyramidal) neurons have sometimes 10K inputs. They get reduce to a continuous response at the soma. Hence for every space reduction you do, there must be a neuron who does it with its dendrites. You are basically looking at how single neurons compute, by managing a manifold for themselves. Hence this doesn't contradict with the single cell research, this is a differential analysis tool for the single neuron research. The output mapping is how single neurons fire (frequency type etc).all via the soma and proteins on the membrane.