Untrained CNNs Exceed Backpropagation in V1 Alignment at High Evaluation Resolution: A Systematic RSA Comparison of Four Learning Rules Against Human fMRI

The gist: Early visual alignment is architecturedriven, learning rules differentiate only at intermediate areas, and all rules converge at the highest levels of the hierarchy<ref:2604.16875#pg2>

Systematic Comparison of Learning Rules

This study presents 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 using Representational Similarity Analysis (RSA)<ref:2604.16875#pg4> The 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<ref:2604.16875#pg2> All models process stimuli at 224 × 224 resolution; results are averaged across 5 random seeds<ref:2604.16875#pg4> Crucially, we include an untrained random-weights baseline that reveals the dominant role of architecture<ref:2604.16875#pg2> At V1/V2, the untrained baseline exceeds backpropagation (ρ = 0.076 vs. ρ = 0.034; ∆ρ = +0.044, p < 0.001), and STDP achieves the highest V1 alignment among trained rules (ρ = 0.064) At LOC, only BP reliably exceeds the random baseline (ρ = 0.012 vs. −0.005, p < 0.001) At IT, all five conditions converge (ρ = 0.008–0.014) with no significant pairwise differences among trained rules (p > 0.05, FDRcorrected) FA consistently produces the lowest alignment at V1, V2, and LOC (ρ = 0.012 at V1, below all other conditions) These results demonstrate that early visual alignment is architecturedriven, learning rules differentiate only at intermediate areas, and all rules converge at the highest levels of the hierarchy

Architecture Dominance and Layer-wise Hierarchy

The study utilized an identical CNN architecture consisting of three convolutional layers (Conv1–Conv3, 5 Conditions: Random · BP · FA · PC · STDP) and two fully connected layers (FC1, FC2) The layer-to-ROI mapping follows the standard ventral stream correspondence: Conv1→V1, Conv1→V2, Conv3→LOC, FC1→IT A key pattern observed is that all five conditions exhibit a sharp decline in alignment from Conv1/Conv2 to Conv3 At FC1, all conditions recover to similar levels (ρ = 0.008–0.014), with no rule clearly dominating This confirms that convolutional inductive biases drive early alignment, while later layers converge across all learning strategies The random baseline shows the most distinctive pattern: highest at Conv1/Conv2 (exceeding all trained rules), dropping sharply at Conv3 (ρ = −0.005), and partially recovering at FC1 (ρ = 0.008)

Performance of Specific Learning Rules

The performance metrics across the five learning rules show distinct advantages in different visual areas At V1/V2, PC (ρ = 0.056) and STDP (ρ = 0.064) outperform BP (ρ = 0.034) at V1 (both p < 0.001) However, this advantage disappears at LOC and IT, where all rules converge Only BP reliably exceeds the random baseline at LOC (ρ = 0.012 vs. −0.005; ∆ρ = +0.017, p < 0.001) Furthermore, FA consistently produces the lowest brain alignment at V1, V2, and LOC (ρ = 0.012 at V1, below all other conditions) This suggests that random feedback filters actively disrupt the representational structure that convolutional architecture provides for free

Convergence and Limitations

At IT, all five conditions produce similar alignment (ρ = 0.008–0.014), with no pairwise comparison among trained rules reaching FDR significance (p > 0.05 for all) This convergence indicates that at the level of abstract categorical representations, no learning rule produces systematically more brain-like structure than any other The paper notes that the small CNN architecture limits representational capacity and may explain the IT convergence A limitation mentioned is that STDP was applied to static images via Poisson spike trains, discarding the temporal dynamics that STDP was designed to exploit Additionally, the best-layer-per-ROI analysis reveals that FC1 sometimes produces higher V1 alignment than Conv1, complicating the anatomical mapping assumption

Implications for Computational Neuroscience and Machine Learning

The findings carry two concrete methodological messages for computational neuroscience and machine learning For computational neuroscience, studies comparing trained models to brain data must always include an untrained baseline with the same architecture, as this control can be essentially complete at V1 For machine learning, the inductive biases of the convolutional architecture (local connectivity, weight sharing, ReLU, pooling) are more important than the weight-update rule for early-layer feature representations When designing architectures for brain-inspired AI, the structural priors deserve as much attention as the learning algorithm The dissociation between task performance and brain alignment suggests that at least for small CNNs trained on CIFAR-10, the relevant axis is not “how well does it classify” but “what representational structure does the objective induce at each layer” These results suggest that, at least at this scale, the search for biologically plausible learning rules should focus on preserving architectural inductive biases at early areas, rather than optimizing for task performance The paper concludes that architecture dominates early visual representations (V1/V2), while higher areas converge regardless of the learning rule

Conclusion

In our small-scale setup, the learning rule has little effect on cortical alignment: architecture dominates early visual representations (V1/V2), where the untrained baseline exceeds all trained rules, and all rules converge at higher areas (IT), with no reliable pairwise differences—though this convergence may partly reflect model capacity limitations Only LOC shows a selective learning effect, where BP alone reliably exceeds the random baseline Among trained rules, PC and STDP preserve more V1-like structure than BP, suggesting that local unsupervised learning retains architectural inductive biases that global error signals may disrupt FA consistently produces the lowest alignment despite meaningful task accuracy These results suggest that, at least at this scale, the search for biologically plausible learning rules should focus on preserving architectural inductive biases at early areas, rather than optimizing for task performance The paper concludes that architecture dominates early visual representations (V1/V2), while higher areas converge regardless of the learning rule The gist.

1. textbfReweighting Learning Rules for Early Feature Preservation in V1/V2: Implementing STDP or PC over BP for Low-Level Representations. This improvement involves replacing backpropagation with Spike-Timing-Dependent Plasticity (STDP) or Predictive Coding (PC) when training CNNs, as the paper notes that STDP and PC lead among trained rules at V1/V2, suggesting these local updates preserve more V1-like structure than BP. This system would be optimized to maintain representations closer to the architectural inductive biases during early processing stages.

2. textbfArchitecture-Driven Feature Optimization via Architectural Priors: Prioritizing CNN Structure over Weight Update Rules. The improved AI system should focus on leveraging the convolutional inductive biases (local connectivity, weight sharing, ReLU, pooling) as a primary driver for feature learning rather than relying solely on backpropagation. This ensures that early visual alignment is architecturedriven, meaning the network structure itself dictates how representations align with human visual cortex at V1/V2.

3. textbfContext-Aware Learning Rule Selection Based on Brain-Score Potential: Dynamically choosing the optimal learning rule based on the target brain region (ROI). The system would assess which rule best aligns with local fMRI data; for instance, using STDP or PC at V1/V2 to maximize V1-like structure while perhaps relying on BP only at LOC where it reliably exceeds the random baseline. This allows the AI to adapt its learning strategy based on where it needs to match known brain hierarchies.

4. textbfRobustness Against Training Artifacts: Integrating an Untrained Baseline for Architectural Control. The improved system must include a mandatory untrained random-weights baseline during evaluation, as the paper demonstrates that this exceeds backpropagation at V1 and reveals the dominant role of architecture. This prevents the confusion between learning effects and architectural effects when assessing representation quality.

5. textbfConvergence Monitoring for High-Level Abstraction: Identifying IT Convergence as a Capacity Limit Indicator. The system should monitor alignment across all five rules, specifically noting that all five conditions converge at the IT area, suggesting this convergence might be due to a capacity limitation of the small CNN, allowing researchers to infer limits on representational power based on convergence behavior.

Revised version (v4): noise ceiling replaced, analyses restricted to stimulus pairs presented in different fMRI runs, repaired evaluation of all five conditions. We compare four learning rules (backpropagation (BP), feedback alignment (FA), predictive coding (PC), and spike-timing-dependent plasticity (STDP)) in identical convolutional networks against human fMRI from THINGS-fMRI (720 stimuli, 3 subjects) using representational similarity analysis (RSA), together with an untrained random-weights baseline that isolates the contribution of architecture. Models are trained on 32 px CIFAR-10 and evaluated at 224 px; results are averaged across 5 seeds, with uncertainty from resampling stimuli. At V1, the untrained baseline exceeds BP (ρ= 0.053 vs. 0.022; Δρ= +0.031, 95% CI [0.019, 0.043]), and the same holds at V2. This advantage depends on the evaluation resolution: in the 5-seed model set of the resolution study (gap +0.030 at 224 px), it vanishes when the models are evaluated at their 32 px training resolution (Δρ= -0.001, 95% CI [-0.011, 0.009]; arXiv:2608.12408). At LOC, BP, PC and STDP exceed the untrained baseline, FA is borderline, and the trained rules do not differ from each other. At IT, no two conditions can be distinguished at the low reliability of the data (leave-one-subject-out lower bound 0.021). FA has the lowest alignment at V1. At V1, an untrained network thus matches BP at the training resolution and exceeds it at 224 px; learning-rule differences are resolvable only at an intermediate area.