Untrained CNNs Exceed Backpropagation in V1 Alignment at High Evaluation Resolution: A Systematic RSA Comparison of Four Learning Rules Against Human fMRI
The study compared four learning rules—backpropagation, feedback alignment, predictive coding, and STDP—on identical convolutional neural networks against human visual cortex data using RSA. Results show early visual representation alignment is dominated by the network's architecture in areas like V1/V2. Learning rules only show differences at intermediate layers; all methods converge at higher levels (IT), meaning architecture matters more than the learning rule for early feature extraction.