brain-alignment-learning-rules-comparison

name	brain-alignment-learning-rules-comparison
description	Comparative methodology for brain alignment across learning rules (BP, FA, PC, STDP). Key finding: single training epoch reduces V1 alignment by 25-90%. BP most destructive, PC and STDP preserve brain-like structure. Use when: brain alignment, representational similarity analysis, biologically plausible learning, visual cortex modeling, learning rule comparison. arXiv: 2605.30556

Brain Alignment Across Learning Rules

Paper: Supervised Training Rapidly Degrades Early Visual Cortex Alignment Across Biologically Plausible Learning Rules
arXiv: 2605.30556
Authors: Nils Leutenegger
Category: cs.LG; q-bio.NC
Published: 2026-05-28

Core Findings

The Paradox

Random, untrained neural networks consistently match or exceed trained networks in representational similarity to early visual cortex (V1). This challenges the assumption that learning improves brain alignment.

Key Results

Metric	Backpropagation	Feedback Alignment	Predictive Coding	STDP
V1 alignment drop (Δr)	-0.080	Moderate	~-0.04	~-0.04
Preservation of brain-like structure	Worst	Moderate	Best	Best
LOC alignment increase	Largest	Smaller	Small	Small

1. Single epoch of training reduces V1 alignment by 25-90%, depending on learning rule
2. Backpropagation reduces V1 alignment most severely (Δr = -0.080)
3. Predictive Coding and STDP preserve substantially more brain-like structure (Δr ~ -0.04)
4. Object-selective cortex (LOC) shows opposite (but weaker) tendency — BP increases alignment during training

Why This Happens

• Untrained architectures capture low-level visual statistics through inductive biases alone (architecture, connectivity patterns)
• Global error signals (BP) reshape early representations more aggressively
• Local learning rules (PC, STDP) better preserve brain-like structure because they operate locally without global optimization pressure

Reusable Patterns

Pattern 1: Inductive Bias First, Training Second

• Untrained networks encode meaningful visual statistics via architecture alone
• Design architectures with strong inductive biases before adding learning
• For brain-aligned models, prioritize architectural constraints over training objectives

Pattern 2: Local vs Global Learning Rule Selection

• Use local learning rules (PC, STDP) when brain alignment is the goal
• Use backpropagation when task performance is the goal (accepts V1 misalignment)
• Consider feedback alignment as a middle ground

Pattern 3: Multi-ROI Alignment Tracking

• Different brain regions respond differently to training
• V1 (early visual): alignment decreases with training
• LOC (object-selective): alignment may increase with training
• Track multiple ROIs simultaneously for comprehensive alignment assessment

Pattern 4: RSA-Based Brain Alignment Measurement

• Use Representational Similarity Analysis (RSA) with Spearman correlations
• Compare model RDMs (Representational Dissimilarity Matrices) to brain RDMs
• Use standardized stimulus sets (e.g., THINGS database, 720 object images)
• Measure at multiple training checkpoints for temporal dynamics

Implementation Guidance

1. For brain-aligned V1 models: Start with untrained architecture, apply minimal local learning
2. For task-oriented models with brain alignment: Use PC or STDP instead of BP for early layers
3. For evaluation: Track RSA alignment at 8+ checkpoints (epochs 0-40 minimum)
4. For stimulus design: Use diverse object databases (720+ images) with multiple subjects

Connections to Existing Skills

• predictive-coding-light: PC learning rules — this paper validates PC for brain alignment
• untrained-cnns-match-backprop-v1: Untrained CNNs match backprop at V1 — direct confirmation
• decoding-encoding-alignment-critique: Brain-model alignment critique — complementary perspective
• vlm-visual-cortex-alignment-robustness: VLM visual cortex alignment — extends to training dynamics

Pitfalls

• Do NOT assume training always improves brain alignment — it often degrades it in early visual areas
• The LOC trend is weaker than the V1 trend — do not over-interpret higher-level alignment improvements
• Results are specific to the THINGS database and 3 subjects — may not generalize to all stimuli/subjects
• The absolute alignment values matter more than relative changes — small Δr may still be significant