Neuroscience

Bidirectional representational alignment between biological and artificial neural networks

Samuel Kostousov, Abhinn Kaushik, Brokoslaw Laschowski

Featured August 23, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

By carefully shaping how artificial intelligence models organize their internal information, like adjusting a spectral "fingerprint", the authors made them much better at both understanding and being understood by real brain activity.

In depth

The paper introduces a computational framework to systematically steer representational geometry in artificial neural networks during training. By integrating spectral regularization with contrastive learning, the authors demonstrate that manipulating the spectral decay exponent of learned representations can significantly improve the bidirectional alignment between artificial and biological neural networks, primarily by increasing reverse predictivity.

Key Takeaways

1
The paper demonstrates that spectral regularization can systematically steer the representational geometry of artificial neural networks.
2
Steering representational geometry, specifically increasing the spectral exponent, substantially improves reverse predictivity and overall bidirectional alignment.
3
Improved bidirectional alignment is accompanied by a reorganization of the shared representational subspace and reduced effective dimensionality.

Conceptual Flow

HIGH LEVEL

Shaping AI's Internal Logic

The authors taught an AI model to learn about images while also making sure its internal thinking patterns matched a specific mathematical shape.

Image Data

Brain Activity

Train AI with Special Rule

AI's Shaped Thinking

AI Thinks More Like a Brain

By shaping the AI's thinking, it became much better at predicting brain activity, and the brain also became better at predicting the AI's thoughts.

AI's Shaped Thinking

Brain Activity

Compare How They Match

Better Two-Way Understanding

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.