learning-rules-brain-alignment-comparison

name	learning-rules-brain-alignment-comparison
description	Comparative methodology for analyzing brain alignment across learning rules (backpropagation, feedback alignment, predictive coding, STDP). Tracks representational similarity analysis (RSA) alignment to human fMRI data during training. Key finding: local learning rules (PC, STDP) preserve brain-like structure better than global error signals (BP). Use when: (1) analyzing learning rule effects on neural representations, (2) understanding why untrained networks match brain activity, (3) designing biologically plausible training methods, (4) comparing representational dynamics across learning paradigms. Activation: brain alignment, RSA, learning rules comparison, backprop vs biological learning, V1 cortex, fMRI alignment, STDP, predictive coding, representational similarity.
license	Complete terms in LICENSE.txt
metadata	{"arxiv_id":"2605.30556","published":"2026-06-01","authors":"Author names from paper","tags":["neuroscience","brain-alignment","learning-rules","RSA","fMRI","backpropagation","predictive-coding","STDP","V1-cortex"]}

Learning Rules Brain Alignment Comparison

Comparative framework analyzing how different learning rules affect representational similarity to human visual cortex during training.

Background

Random, untrained neural networks consistently match or exceed trained networks in representational similarity to early visual cortex. This paradox challenges assumptions about learning improving brain alignment. This methodology systematically compares four learning rules to understand this phenomenon.

Core Finding

: A single epoch of training reduces V1 alignment by 25-90%, depending on the learning rule: