Introduction Rodents can discriminate object shape, size, and texture through active whisker contacts. Accordingly, rodent primary somatosensory cortex (S1) contains topographic representations of the whiskers, and neurons in S1 represent multiple stimulus features associated with whisker based touch. Although these representations have been characterized experimentally, the computational mechanisms that generate them remain unclear. In addition, there is substantial convergence between whisker-based touch and other sensory modalities in multiple cortical areas. We test whether training whisker and visual input representations to align in a self-supervised way can reproduce representations in barrel cortex.
Methods Using a morphologically accurate model of the mouse whisker array [2], we simulated whisker contact sequences in terms of the radial distance and angle of contact. These signals were encoded using per-whisker temporal units; these units also integrated the signals over the whisk cycle. Learned whisker embeddings were then arranged into a 5×7 topographic grid approximating barrel cortex. The encoder was trained using a contrastive objective that aligned whisker embeddings with visual embeddings produced by MouseNet [1]. Condition-averaged embeddings were used to construct representational dissimilarity matrices (RDMs), which were compared to neural population RDMs using Pearson correlation and established noise correction procedures.
Results The anatomically structured whisker network produced embeddings aligned with neural population geometry. Representational similarity analysis (RSA) yielded a raw RSA of 0.54 between model and neural dissimilarity matrices across stimulus conditions, with a noise-corrected estimate of 1.14 relative to the neural reliability ceiling. Alignment was reduced when temporal integration was removed or when whisker embeddings were not arranged in a topographic grid, indicating that both recurrent dynamics and barrel-like spatial organization contribute to emergent representational structure.
Discussion These results suggest that biologically relevant representations may arise from temporally integrated whisker signals constrained by anatomical hierarchy, even when the model is trained using a multimodal contrastive objective rather than explicit supervision on the whisker discrimination task, as in [3]. The dependence of alignment on both recurrent dynamics and topographic organization supports the hypothesis that barrel organization and recurrent dynamics are key determinants of tactile shape coding. This framework provides a principled foundation for extending biologically constrained models toward multimodal cortical integration.
References 1. Shi, J., Tripp, B., Shea-Brown, E., Mihalas, S., & A. Buice, M. (2022). MouseNet: A biologically constrained convolutional neural network model for the Mouse Visual Cortex. PLOS Computational Biology, 18(9). https://doi.org/10.1371/journal.pcbi.1010427 2. Bresee, C. S., Belli, H. M., Luo, Y., & Hartmann, M. J. Z. (2023). Comparative morphology of the whiskers and faces of mice (Mus musculus) and rats (Rattus norvegicus). Journal of Experimental Biology, 226(19), jeb245597. https://doi.org/10.1242/jeb.245597 3. Chung, T., Shen, Y., Kong, N. C. L., & Nayebi, A. (2025). Task-optimized convolutional recurrent networks align with tactile processing in the rodent brain. arXiv. https://doi.org/10.48550/arXiv.2505.18361
Acknowledgement CN was supported by NSERC RGPIN-2025-04919 and Alliance International Catalyst. KK and MJH were supported by NIH award R01 NS-116277.
