Supervised Training Rapidly Degrades Early Visual Cortex Alignment Across Biologically Plausible Learning Rules
A new research paper explores how supervised training impacts the alignment of artificial neural networks with the human visual cortex. The study found that even a single epoch of training can significantly degrade this alignment, particularly in early visual areas like V1. Different learning rules showed varying effects, with backpropagation causing the most severe degradation, while predictive coding and spike-timing-dependent plasticity preserved more brain-like structure. AI
IMPACT Suggests that current training methods may actively move AI representations away from biological plausibility, prompting a re-evaluation of learning rule efficacy.