Emergent Geometry in Neural Representations of the Visual World
Keywords: Neuroscience, Vision, Geometry, Neural Manifolds, Deep Learning, Convolutional Neural Networks, Neural Representations
Abstract: How does the brain transform the complex visual world into meaningful representations that facilitate generalization across diverse conditions? One hypothesis is that the geometric structure of neural manifolds mirrors causal structures in the environment, facilitating strong generalization across natural contexts. The analysis of neural manifold structure has yielded neuroscientific insights in domains such as navigation and motor control, which often possess simple, low-dimensional structure. Vision, however, presents unique challenges due to its more complex, high-dimensional, hierarchical structure. Leveraging a digital twin model of primate V4 neurons, we conduct targeted in silico experiments that allow us to systematically investigate the relationship between the structure of the visual world and its encoding in the visual cortex. Our findings reveal structural equivalences between world properties and neural activity patterns for rotating objects and textures. Specifically, we demonstrate the emergence of equivariant representations that disentangle latent factors such as object identity and orientation. Finally, we demonstrate that these representations facilitate out-of-distribution generalization, as a decoder trained to linearly decode the orientation of one texture can successfully transfer to novel textures. Remarkably, artificial neural networks trained on object recognition tasks exhibit similar geometric principles. These results provide empirical support for the mirroring hypothesis in visual processing and suggest universal principles govern the formation of neural representations across biological and artificial vision.
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13417
Loading