Backpropagation Destroys V1 Brain Alignment in One Epoch; Biologically Plausible Alternatives Preserve It
Summary
Third paper in a series finds standard backprop destroys V1 fMRI alignment within one training epoch; feedback alignment, predictive coding, and STDP preserve it on the THINGS dataset.
Originally reported by reddit.com
Read the original article →Original headline: Backpropagation Destroys V1 Brain Alignment in One Epoch; Biologically Plausible Alternatives Preserve It