#27 · ML / AI methods

Local learning and low-rank subspaces explain model-brain alignment

The dominant thread is mechanistic accounting for model-brain alignment. Two preprints show that predictive-coding and manifold-capacity networks trained with local learning rules match or exceed prior models in predicting ventral-stream responses, with the Layerwise Multiscale Manifold Capacity Maximization work matching baselines on macaque data. Separately, a preprint locates LLM-brain linear alignment during naturalistic listening in a low-rank subspace occupying 1-3% of embedding dimensions, assessed with ridge reduced-rank regression on ECoG and fMRI. A published study uses ECoG variance partitioning to show that speech-model alignment largely reflects interpretable features, and that adding them improves encoding. A neuronal-identity preprint is a related but distinct axis, proposing latent groupings from molecular, morphological and electrophysiological profiles to model transferable population dynamics. What is new is the shift from benchmark alignment scores toward identifying the biologically or interpretably meaningful components driving that alignment.