Much recent neuroscience has focused on the geometry of population codes, asking what information a group of neurons carries using measures that ignore exactly which neurons encode which features. The new work asks whether this indifference is justified. The team developed methods to test whether neural tuning is aligned across different brains and across deep convolutional neural networks trained on natural inputs. If tuning directions were arbitrary, alignment between different systems should be no better than chance; if some axes are systematically privileged, alignment should appear repeatedly.

That is what the researchers found. Across both vision and audition, brains and deep convolutional networks favoured particular axes for representing the natural world. Moreover, the representational axes of networks trained on natural inputs were aligned with those observed in perceptual cortices.

The alignment also has practical consequences for computational neuroscience. Axis-sensitive model-to-brain similarity metrics, the authors report, better differentiated competing models of biological sensory systems than measures blind to tuning axes. This means that how a model represents information—not just how much information it captures—can determine whether it is judged similar to a given brain region, sharpening a widely used tool for evaluating mechanistic models of the sensory cortex.

The study goes on to argue that privileging certain axes is computationally useful rather than incidental. According to the paper, coding schemes that favour particular axes can facilitate task performance, reduce wiring costs for downstream circuits, and improve generalization to new inputs.

The reported findings span multiple levels of sensory representation. The authors present evidence that privileged axes extend beyond category selectivity and reach higher-order cortical representations, indicating that the phenomenon is not confined to low-level sensory features. Figures in the paper document sensory cortex alignment, model-to-model and model-to-brain alignment, and the computational implications of axis privileging, alongside analyses of what the preferred axes actually are in brains and networks.

The work is published as subscription content in Nature Human Behaviour, and the detailed quantitative results sit behind the paywall. Nonetheless, the abstract-level claims outline a reframing of how neural tuning should be studied: not merely as a flexible population code whose geometry carries the message, but as a structured scheme in which the identities of the tuning axes themselves carry information about the constraints shaping a system.

Because the reported results are summarized at the abstract level and the full quantitative analyses are not freely available, independent verification of the effect sizes and their scope will depend on access to the complete paper and on replication by other laboratories. The findings as presented, however, establish that the specific axes along which neurons tune are not interchangeable, and that this structure is shared, at least in part, between perceptual cortices and deep networks trained on natural inputs.

Subject of Research: Alignment of privileged neural representational axes between brains and artificial neural networks

Article Title: Privileged representational axes in biological and artificial neural networks

Article References: Khosla, M., Srivastava, S., Williams, A. H., McDermott, J., & Kanwisher, N. (2026). Privileged representational axes in biological and artificial neural networks. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02515-3

Image Credits: AI Generated

DOI: 10.1038/s41562-026-02515-3

Keywords: neural coding, representational axes, deep neural networks, sensory cortex, vision, audition, population codes, model-brain similarity, neural tuning, Nature Human Behaviour, Privileged, representational