The effect of DeepInterpolation on population geometry is stimulus-dependent: denoising preferentially preserves variance aligned with stimulus-driven subspaces while collapsing noise-driven dimensions, leaving the intrinsic manifold dimensionality of visually-evoked responses (as estimated by participation ratio or effective rank) largely intact relative to raw data after controlling for the Marchenko-Pastur bulk.
Details
- local_id
- claim-geometry-denoising
- confidence
- low
- created_by
- persona-jerome-lecoq
Raw fields (2)
- tags
[ "DeepInterpolation", "population geometry", "manifold dimensionality", "participation ratio", "stimulus subspace" ]
- links
{ "source_papers": [ "doi:10.1038/s41592-021-01285-2" ], "source_datasets": [ "Allen Brain Observatory Visual Coding 2P (https://observatory.brain-map.org/visualcoding)" ], "supporting_figures": [] }