Meenakshi Khosla¹ (), Sudhanshu Srivastava¹, Alex Williams³, Josh McDermott², Nancy Kanwisher²; ¹University of California San Diego, ²Massachusetts Institute of Technology, ³New York University

How is information coded in the brain? To answer this question, neuroscientists have increasingly adopted methods such as representational similarity analysis, linear encoding and decoding, canonical correlation analysis and centered kernel alignment that analyze the geometry of the population code while disregarding the actual tuning of neurons. But are neural tunings in fact arbitrary and irrelevant, or might they matter, privileging some representational axes over others? We developed methods to probe for privileged representational axes in biological and artificial neural networks, and applied them to multiple types of neural data from diverse brain systems, and to DCNNs trained on natural sensory stimuli. We found that representational axes were consistent between individuals, between artificial neural networks varying in architecture and learning objectives, and between brain systems and artificial neural networks trained on the same modality. These results indicate that representational axes in neural systems are not arbitrary, and can arise in artificial systems with none of the priors and meanings assigned to these axes by humans. We further found that the privileged axes used in the brain and DCNNs confer important computational advantages, including economy in the number of active neurons, minimization of downstream wiring costs and improved generalization under biologically realistic constraints. Finally, our metrics of axis alignment also distinguish the fit of models to the brain that are not well discriminated by standard metrics. These findings underscore the importance of representational axes in both biological and artificial neural systems, offering new insights into their origins and functional relevance.