🤖 AI

Bidirectional representational alignment between biological and artificial neural networks

This paper demonstrates that steering the spectral geometry of self-supervised vision models during training can systematically modulate and significantly improve bidirectional representational alignment with biological neural networks, achieving a 55% relative improvement in bidirectional predictivity by reducing effective dimensionality and reorganizing the shared representational subspace.

Original authors: Samuel Kostousov, Abhinn Kaushik, Brokoslaw Laschowski

Published 2026-08-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Samuel Kostousov, Abhinn Kaushik, Brokoslaw Laschowski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

To understand how intelligence works, scientists often look at the brain, the biological machine that thinks, and compare it to artificial neural networks, the computer programs designed to mimic it. For years, researchers have tried to see if these two systems "think" in similar ways by comparing their internal maps of the world. When a computer sees a picture of a cat, it creates a specific pattern of activity; when a monkey sees the same picture, its brain creates a different pattern. Scientists measure how well the computer's pattern predicts the monkey's brain activity, a process known as forward prediction. Recently, a new question emerged: can we also predict the computer's pattern just by looking at the monkey's brain? The answer was surprisingly one-sided. The computer's internal maps were excellent at guessing what the brain would do, but the brain's activity was a poor guesser of what the computer would do. This imbalance suggested that while artificial intelligence had learned to mimic the brain's output, it had not yet learned to mimic the brain's underlying structure.

A team of researchers set out to fix this one-sided relationship by changing the shape of the artificial network's internal maps during its training. They hypothesized that the way information is organized inside the computer—specifically, how the importance of different pieces of information is distributed—might be the missing key. To test this, they built a system that could gently nudge the computer's learning process to organize its information in a specific way, similar to how the brain naturally organizes it. They used a standard vision model, a type of artificial intelligence trained to recognize images without human labels, and applied a new rule during its learning phase. This rule encouraged the network to arrange its internal signals so that a few signals carried most of the weight, while many others carried very little, a pattern that mirrors the natural decay of importance found in biological brains.

The results showed that this simple adjustment had a profound effect. When the researchers steered the computer's internal geometry to match this natural pattern, the ability of the brain's activity to predict the computer's behavior improved dramatically. In fact, the computer's internal maps became much better at being predicted by the brain, closing the gap that had existed before. While the computer's ability to predict the brain's activity dropped slightly, the massive gain in the reverse direction meant that the two systems became much more balanced. The overall alignment between the biological and artificial networks improved by fifty-five percent relative to the original setup. This improvement was not just a statistical fluke; it came with a clear change in the computer's structure. The network began to rely on fewer, more efficient dimensions to store information, and the parts of the network that were most similar to the brain's activity became almost perfectly symmetrical in their relationship.

The study suggests that the way information is organized inside a learning system is just as important as the information itself. By simply changing the rules of how the computer distributes its attention across different signals, the researchers made the artificial network's internal world look much more like the biological one. This does not mean the computer has suddenly become conscious or that it has solved the mystery of human thought. Rather, it demonstrates that the geometry of learned representations—the shape of the internal maps—is a property that can be controlled to bring artificial and biological systems closer together. The findings indicate that the previous imbalance was not a permanent flaw in artificial intelligence but a consequence of how these networks were allowed to organize themselves. By guiding that organization, scientists can create models that are not only better at solving tasks but also better mirrors of the biological systems they aim to understand.

Technical Summary: Bidirectional Representational Alignment Between Biological and Artificial Neural Networks

Problem Statement
Recent research in computational neuroscience and machine learning has established that representational alignment between biological and artificial neural networks is inherently asymmetric. Specifically, artificial model representations predict biological neural responses (forward predictivity) significantly better than biological responses predict model representations (reverse predictivity). While task optimization can recover fundamental principles of neural representation, improvements in task performance do not always correlate with improved alignment. This asymmetry raises a critical question: does the representational geometry of learned features contribute to this bidirectional misalignment? The authors hypothesize that the spectral geometry of representations can be systematically steered during training to modulate this alignment.

Methodology
To test this hypothesis, the authors developed a computational framework that integrates spectral regularization with bidirectional predictivity analyses.

Model and Dataset: The framework was implemented using self-supervised contrastive vision models (SimCLR with a ResNet-50 encoder). Training utilized a subset of ImageNet-1k, while evaluation employed a macaque inferotemporal (IT) cortex benchmark comprising neural responses to 3,200 visual stimuli.
Spectral Steering: The core innovation involves adding a spectral regularization term to the standard InfoNCE contrastive loss. This term penalizes deviations from a target spectral decay exponent ( $\hat{\alpha}$ $\overset{α}{^}$ ) in the eigenspectrum of the learned representations.
- The regularization is applied to the activations of encoder.layer4.0.bn1, a layer known for strong forward predictivity.
- The spectral exponent $\alpha$ is estimated via ordinary least squares on the log-log spectrum of the normalized explained variance (principal components 5–15).
- The total loss function is $L_{total} = L_{InfoNCE} + \lambda|\alpha - \hat{\alpha}|$ , where $\lambda$ controls the regularization strength.
Evaluation Metrics:
- Forward Predictivity: Measures how well model activations predict neural responses (ridge regression from model to neuron).
- Reverse Predictivity: Measures how well neural responses predict model activations (ridge regression from neuron to model).
- Bidirectional Predictivity (BPI): A composite metric ( $BPI = \frac{2R_{fwd}R_{rev}}{R_{fwd} + R_{rev}}$ ) that quantifies alignment while penalizing asymmetry.
- Effective Dimensionality: Calculated as the participation ratio of the explained variance spectrum.
Experimental Design: Models were trained with varying target spectral exponents ( $\hat{\alpha}$ ) and regularization strengths ( $\lambda$ ). Analysis was performed based on the measured spectral exponent ( $\alpha$ ) rather than the target, to account for cases where regularization strength did not achieve the exact target.

Key Results
The experiments demonstrated that steering representational geometry systematically influences bidirectional alignment:

Control of Spectral Geometry: Spectral regularization successfully steered the measured spectral exponent of the learned representations. Stronger regularization ( $\lambda$ ) produced closer agreement between the measured and target exponents.
Asymmetry Reduction:
- Reverse Predictivity: Increasing the measured spectral exponent ( $\alpha$ ) substantially increased reverse predictivity. At $\lambda=3$ , reverse explained variance improved from 12% (baseline) to 29% in the highest $\alpha$ bin.
- Forward Predictivity: Forward predictivity showed a modest decrease under spectral regularization compared to the baseline but remained relatively stable across different measured $\alpha$ values at a fixed regularization strength.
- Bidirectional Predictivity: The net effect was a significant improvement in bidirectional predictivity, increasing from approximately 20% to 31% (a 55% relative improvement). This indicates a substantial reduction in the forward-reverse asymmetry.
Mechanistic Insights:
- Common vs. Unique Units: The improvements were most pronounced in "common units" (the top 20% of model units ranked by reverse predictivity). In this subset, forward and reverse predictivity became approximately symmetric at intermediate spectral exponents.
- Dimensionality: Increased spectral exponents were accompanied by a reduction in effective dimensionality, suggesting that steering geometry concentrates representational variance into fewer effective dimensions.
- Subspace Reorganization: The findings suggest that spectral regularization preferentially reorganizes the representational subspace shared between artificial and biological networks, improving the linear recoverability of neural activity from model representations.

Significance and Claims
The paper provides a controlled framework demonstrating that representational geometry is a steerable property that directly modulates bidirectional representational alignment.

Challenging Fixed Asymmetry: The authors refute the notion that the forward-reverse asymmetry is a fixed property of task-optimized models. Instead, they show it can be systematically reduced by altering the spectral geometry during learning.
Geometry as a Causal Factor: The results support the hypothesis that representational geometry contributes to alignment quality. Specifically, increasing the spectral exponent improves the "reverse" direction of alignment without catastrophic loss in the "forward" direction.
Beyond Matching Exponents: The study notes that alignment continued to improve even beyond the spectral exponent ( $\alpha \approx 1$ ) typically observed in the visual cortex, suggesting that simply matching biological spectral statistics is not sufficient to maximize alignment; rather, the steering process itself induces beneficial reorganizations.
Limitations and Scope: The authors modestly frame these findings as an initial demonstration using a specific architecture (ResNet-50) and a single neural benchmark. They acknowledge that the spectral exponent is only one characterization of geometry and that other geometric properties may also play a role. Future work is needed to generalize these findings across different architectures, learning objectives, and datasets.

In summary, this research establishes a method to "steer" the internal geometry of artificial networks to better align with biological systems, offering a new avenue for mechanistic interpretability and the study of intelligence emergence.

Technical Summary: Bidirectional Representational Alignment Between Biological and Artificial Neural Networks

More like this