STDP and Predictive Coding best match primate vision across human and macaque brains

A new study systematically compared five learning rules—backpropagation (BP), feedback alignment (FA), predictive coding (PC), spike-timing-dependent plasticity (STDP), and an untrained baseline—against neural recordings from both humans (fMRI) and macaques (single-unit electrophysiology for V1/V2, multi-electrode for V4/IT). The same model weights were used across conditions. In early visual areas, STDP (ρ≈0.30) and PC (ρ≈0.28) outperformed other rules in macaque V1/V2, mirroring their position in human V1. This confirms the pattern is not an fMRI artifact and is conserved across species. Notably, the untrained baseline appeared better than BP in human fMRI (low SNR), but in macaque electrophysiology (higher SNR), STDP and PC clearly surpassed Random, resolving that ambiguity.

In higher visual areas (IT), alignment with brain data scaled with model capacity, not learning rule. A ResNet-50 pretrained on ImageNet achieved ρ≈0.25 at macaque IT, while a small 3-conv CNN scored only 0.07–0.14 across all rules, suggesting a capacity floor. Cross-species IT rankings were statistically insignificant (Kendall's τ=0.00, p=1.00, n=5), meaning the small sample size makes any conclusion impossible. A confound exists: V1/V2 used texture stimuli while V4/IT used objects, and stimulus control showed IT rankings weakly inverted across sets (τ=−0.40). Thus, observed cross-species IT differences may be partially stimulus-driven. The study underscores the importance of capacity control and stimulus matching in linking learning rules to brain data.