Cross-Species Study Reveals AI Matches Early Vision, Splits on Higher Areas

A new study from Nils Leutenegger extends prior work on learning rules and brain alignment to a cross-species comparison. Using representational similarity analysis (RSA), the author tested five learning rules—backpropagation (BP), feedback alignment (FA), predictive coding (PC), spike-timing-dependent plasticity (STDP), and a random-weights baseline—against macaque electrophysiology data (MajajHong2015 for V4/IT and FreemanZiemba2013 for V1/V2). The same model weights from earlier human fMRI experiments were used. Results show that early visual cortex (V1/V2) alignment is robust across species, with all models achieving higher rho values in macaques (0.15–0.30) than in humans (0.01–0.08), likely due to electrophysiology's higher signal-to-noise ratio. STDP and PC again led among trained rules, consistent with human V1 rankings.

However, higher-area alignment (IT) tells a different story. Learning rule rankings showed no detectable correlation across species (Kendall's tau = 0.00, p = 1.00), though the small sample size (n=5) limits statistical power. A pretrained ResNet-50 (ImageNet) achieved rho=0.25 at macaque IT, substantially above all custom CNN conditions (rho=0.07–0.14). This suggests that IT alignment is primarily limited by model capacity and training data, not the learning rule. The study also reports noise ceilings, multi-seed variability (5 seeds), and stimulus-control analyses. These findings demonstrate that early visual alignment generalizes across species, while higher-area alignment depends more on model architecture and domain than on the learning algorithm itself.