When Neural Networks Learn Like Brains: Locality Beats Weight Sharing in Hebbian Vision Models

For decades, neuroscientists and machine-learning researchers have wrestled with a fundamental question: how does the brain learn to see without the powerful but biologically implausible error signals that drive modern artificial intelligence? A new study published in Frontiers in Computational Neuroscience by Patrick Inoue, Florian Röhrbein, and Andreas Knoblauch takes a significant step toward answering this puzzle by systematically dismantling one of the most quietly load-bearing assumptions in biologically inspired visual learning — the idea that neural networks must share their weights across space to perform well.

The research, conducted at Albstadt-Sigmaringen University and Chemnitz University of Technology, tackles a tension that has long lurked beneath the surface of Hebbian convolutional networks. These models, which learn features through local synaptic plasticity rules inspired by the pioneering work of Donald Hebb, have shown impressive performance on visual classification benchmarks. But their success has always been coupled to a specific architectural trick: exact translational weight sharing, where the same filter is copied across every spatial position in an image. While this prior is computationally efficient and statistically powerful, it is not anatomically literal. Real neurons in the visual cortex do not literally share identical synaptic parameters across retinotopic positions, and the new study asks whether useful visual representations can still emerge when this crutch is removed.

To answer this question, the researchers turned to Hebbian principal component analysis, or HPCA, a local learning rule rooted in Oja’s classic formulation of synaptic plasticity. The rule combines a Hebbian potentiation term — synapses strengthen when presynaptic and postsynaptic activity coincide — with a projection-dependent stabilizing correction that prevents unbounded growth. Competition among hidden units is introduced through population-wise divisive normalization, a mechanism inspired by canonical normalization models in sensory neuroscience. The result is a learning rule that extracts principal components from input statistics using only information available at or near the synapse, a constraint that aligns far more closely with biological reality than the global error signals of backpropagation.

The experimental design was elegantly simple. The researchers compared three architectural regimes on CIFAR-10, MNIST, and STL-10. A fully connected network served as a non-spatial baseline. A position-specific locally connected architecture, developed specifically for this study, preserved spatially restricted receptive fields while assigning each unit its own unique synaptic weights — no sharing whatsoever. Finally, HPCA was substituted into two established shared-kernel convolutional Hebbian benchmarks, the modular Hebbian-CNN framework and the original SoftHebb CNN, to test whether the same learning principle remains effective when exact weight sharing is restored.

The results were striking. On CIFAR-10, the position-specific locally connected HPCA model significantly outperformed the fully connected reference, achieving 60.56 percent accuracy compared to 52.74 percent — a difference of 7.82 percentage points that held up under rigorous statistical testing with paired-samples t-tests across ten independently initialized runs. Even more remarkably, when the researchers disabled data augmentation in both conditions, the locally connected advantage persisted at 7.52 percentage points, confirming that the benefit was architectural rather than an artifact of transformation-rich training. A plateau-triggered optimizer-switch variant pushed the best locally connected result to 61.42 percent, placing it in the same range as earlier convolutional HPCA models that relied on shared kernels.

The shared-kernel experiments revealed a different story. When HPCA was embedded in the modular Hebbian-CNN benchmark, it reached 74.51 percent under common preprocessing and 76.57 percent in the SoftHebb architecture under reference preprocessing. These numbers substantially exceeded the locally connected results, confirming that exact weight sharing provides a powerful statistical prior for natural-image tasks. However, the researchers were careful to note that this gap cannot be attributed to weight sharing alone. Depth, hierarchy, preprocessing, competition mechanisms, learning-rate schedules, and readout protocols all co-varied across the configurations, making it impossible to isolate any single factor as the sole cause of the performance difference.

One of the most important contributions of this work is its systematic sensitivity analysis of the convolutional benchmarks themselves. The researchers discovered that preprocessing choices had dramatically different effects depending on the benchmark. In the modular Hebbian-CNN framework, switching to the common HPCA preprocessing pipeline increased accuracy by 4.01 percentage points. In the SoftHebb architecture, the same switch actually reduced accuracy by 0.54 percentage points. This protocol dependence was not specific to the HPCA substitution — even the reproduced SoftHebb reference conditions showed sensitivity to learning-rate and schedule changes, failing to recover the published reference accuracy. The findings suggest that raw accuracy differences between Hebbian models may reflect the entire benchmark configuration rather than the intrinsic quality of any particular learning rule.

To address concerns about parameter count, the researchers conducted a careful control experiment. They reduced the locally connected readout from 2,048 to 208 GELU units, matching the approximately 5.9 million parameter scale of the shared-kernel architectures. The parameter-matched locally connected model achieved 59.16 percent, only about 1.4 percentage points below the standard configuration. Yet the shared-kernel models still outperformed it by 15 to 17 percentage points under matched preprocessing. Because this advantage far exceeds the cost of parameter reduction, the researchers concluded that the performance gap does not arise from differences in total model capacity but from the structural benefits of weight sharing and architectural depth.

Receptive-field visualizations provided additional insight into what the different architectures actually learned. The locally connected HPCA model produced heterogeneous, position-specific filters consisting of small-scale chromatic mixtures, local contrast variations, and patch-like high-frequency structure — with no repeated template across the image, consistent with the absence of weight sharing. In contrast, the shared-kernel HPCA configurations exhibited a clear layer-wise progression from coarse block-like chromatic filters in the first layer to more diffuse color-texture combinations in the second and sparse, distributed high-frequency patterns in the third. This hierarchical organization mirrors the kind of feature composition that makes convolutional networks so effective, and it emerged spontaneously from local Hebbian learning when the shared-kernel prior was available.

The broader implications of this study extend beyond benchmark comparisons. By demonstrating that useful frozen visual representations can be acquired without exact translational weight sharing, the researchers have established a stricter test case for biologically constrained visual learning. The position-specific locally connected HPCA architecture represents a middle ground between fully connected networks, which ignore image structure entirely, and convolutional networks, which impose a strong shared-kernel prior that real neural circuits do not literally implement. The work also raises intriguing questions about how biological systems might achieve convolution-like behavior without explicit parameter tying — perhaps through transformation-rich visual experience, lateral Hebbian mechanisms that promote approximate weight alignment, or structural plasticity that adapts receptive-field topology during development. As the field continues to grapple with the gap between biologically plausible learning and state-of-the-art performance, this study provides both a valuable controlled model and a clear roadmap for future investigations into how local plasticity, spatial connectivity, and architectural inductive biases interact to shape visual representation learning in brains and machines alike.

Subject of Research: Hebbian principal-component learning in biologically constrained visual neural network architectures

Article Title: Locality and weight sharing shape Hebbian principal-component learning in biologically constrained visual models

Article References: Inoue, P., Röhrbein, F., & Knoblauch, A. (2026). Locality and weight sharing shape Hebbian principal-component learning in biologically constrained visual models. Frontiers in Computational Neuroscience, 20, Article 1943064. https://doi.org/10.3389/fncom.2026.1943064

Image Credits: AI Generated

DOI: 10.3389/fncom.2026.1943064

Keywords: Hebbian learning, local plasticity, weight sharing, visual representation learning, receptive fields, convolutional neural networks, biological constraints, principal component analysis, CIFAR-10, computational neuroscience, neural computation, unsupervised learning