Platonic representations in the visual cortex: 4 gaps (www.neuroai.science)

Recent research from Huh, Cheung, Wang, and Isola (2024) introduces the Platonic Representation Hypothesis (PRH), which suggests that neural networks trained on large, diverse datasets converge to similar representations, akin to a universal geometry. This hypothesis has gained traction, influencing studies that align unimodal representations across modalities like text and visual data. However, it remains controversial, with evidence pointing to significant gaps between the representations of artificial neural networks (ANNs) and biological brains, indicating divergence rather than convergence. These divergences, such as the predictive, geometry, linearization, and robustness gaps, suggest that while ANNs are adept at various tasks, they operate under different principles compared to the brain, challenging the assumption that better AI models will mirror human cognitive processes. The implications of the PRH for the AI/ML community are profound, particularly for systems neuroscience. If the hypothesis holds true, it may lead researchers to resign to simply waiting for advanced AI models to replicate brain-like representations. Conversely, should it be proven false, it can invigorate research into understanding the distinct mechanisms of the brain versus ANN performance. As gaps in representation and robustness become clearer, they underscore the need for diverse methodologies in AI development, pushing the envelope for more biologically inspired architectures that harness insights from neuroscience, ultimately shaping AI systems that complement human cognitive abilities rather than replicate them.