untrained-cnns-match-backpropagation-at-v1

# Untrained CNNs Match Backpropagation at V1: A Systematic RSA Comparison of Four Learning Rules Against Human fMRI ## Source - **arXiv:** [2604.16875](https://arxiv.org/abs/2604.16875) - **Authors:** Nils Leutenegger - **Published:** 2026-04-18 (revised 2026-04-29) - **Categories:** cs.LG, q-bio.NC ## Abstract A central question in computational neuroscience is whether the learning rule used to train a neural network determines how well its internal representations align with those of the human visual cortex. We present a systematic comparison of four learning rules — backpropagation (BP), feedback alignment (FA), predictive coding (PC), and spike-timing-dependent plasticity (STDP) — applied to identical convolutional architectures and evaluated against human fMRI data from the THINGS-fMRI dataset (720 stimuli, 3 subjects) using Representational Similarity Analysis (RSA). All models process stimuli at 224×224 resolution; results are averaged across 5 random seeds. Crucially, we include an untrained random-weights baseline. ## Key Findings ### 1. Architecture Dominates at Early Visual Areas (V1/V2) - Untrained random-weights baseline **exceeds** backpropagation at V1/V2 alignment - Untrained: ρ = 0.076 vs. BP: ρ = 0.034 (Δρ = +0.044, p < 0.001) - Architecture alone explains most early visual alignment - Learning rules contribute minimally at early stages ### 2. STDP Achieves Best Alignment Among Trained Rules at V1 - STDP: ρ = 0.064 at V1 — highest among all trained learning rules - Suggests biologically plausible plasticity rules better capture early visual processing ### 3. Learning Rules Only Differentiate at Intermediate Areas (LOC) - At LOC, only BP reliably exceeds the random baseline (ρ = 0.012 vs. -0.005, p < 0.001) - FA consistently produces the lowest alignment at V1, V2, and LOC (ρ = 0.012 at V1) ### 4. All Rules Converge at High-Level Areas (IT) - At IT, all five conditions converge (ρ = 0.008–0.014) - No significant pairwise differences among trained rules (p > 0.05, FDR-corrected) - Suggests task-independent architectural constraints dominate at highest levels ### 5. Robustness Checks - Partial RSA confirms all effects survive pixel-similarity control - Seed variability is small relative to between-rule differences at V1/V2 ## Methodology - **Dataset:** THINGS-fMRI (720 naturalistic stimuli, 3 human subjects) - **Architecture:** Identical CNN across all learning rules - **Learning Rules Compared:** BP, FA, PC, STDP, plus untrained baseline - **Evaluation Metric:** Representational Similarity Analysis (RSA) - **Visual Areas:** V1, V2, LOC, IT - **Statistical Rigor:** FDR correction, 5 random seeds, pixel-similarity controls ## Implications - Early visual alignment is **architecture-driven**, not learning-rule-driven - Biologically plausible rules (STDP) outperform BP at early visual areas - Different learning rules cannot be distinguished by brain alignment at highest levels - Suggests rethinking of brain-DNN comparison benchmarks: control for architectural effects - Untrained baselines are essential for proper attribution of brain alignment ## Activation Keywords - representational-similarity-analysis, learning-rules, brain-alignment, fMRI, visual-cortex, STDP, backpropagation comparison, V1 alignment, computational neuroscience