Sudhanshu Srivastava¹ (), Alex Williams², Josh McDermott³, Nancy Kanwisher³, Meenakshi Khosla¹; ¹UCSD, ²NYU, ³MIT

High-level perceptual cortices exhibit privileged representational axes (Khosla et al., 2023), meaning that even among representations with equivalent information content and geometry, specific neural tunings are systematically favored, consistently across participants. But why? Privileged axes corresponding to selective responses to faces, places, food, etc. could plausibly reflect the biological importance of these categories to humans. But that account is challenged by the fact that ANNs that know nothing about the significance of these categories nonetheless contain class-selective neurons (Prince et al., 2023; Khosla et al., 2023; Zheng et al., 2024). Here we test the hypothesis that privileged axes enhance task performance under structural constraints, such as nonlinear activation functions, which disrupt rotational symmetry in networks. Using a novel loss function, we trained neural networks for CIFAR-10 classification while explicitly maximizing or minimizing alignment with the representational axes of a reference network. Alignment was quantified using SoftMatch (Khosla & Williams, 2023), a metric that considers both population-level and neuron-level representations. We evaluated two architectures—AlexNet and MyrtleCNN—training 180 models (100 for AlexNet, 80 for MyrtleCNN) across layers with alignment objectives. Each model was trained for 20 epochs. For every layer, networks trained to maximize alignment with the corresponding layer of the reference achieved significantly higher CIFAR-10 classification accuracy than those trained to minimize alignment (p < 10⁻⁴ for each layer, independent samples t-tests, df = 18). The performance gap widened for deeper layers. These findings support the hypothesis that privileged axes facilitate better task performance under structural constraints like nonlinear activation functions. They highlight the theoretical plausibility that biological activation functions, akin to ReLUs, impart privileged axes to brains, offering a deeper understanding of why specific tuning functions are consistently observed across neural networks, both biological and artificial.