Cross-Species RSA: Comparing Learning Rules and Brain Visual Cortex Alignment

Using human fMRI data and macaque electrophysiology data, the study analyzed how well various learning rules—BP, PC, STDP, FA—align with the visual cortex through RSA (Representational Similarity Analysis).

The key research findings are as follows:

• Early visual cortex (V1/V2) alignment: This tended to be similarly preserved across species. In particular, STDP (ρ ≈ 0.30) and PC (ρ ≈ 0.28) recorded high alignment scores, showing patterns similar to human V1.
• IT (Inferior Temporal) region alignment: This is more influenced by model capacity than by the learning rule. ResNet-50 (ρ ≈ 0.25), pretrained on ImageNet, showed higher alignment than custom CNNs (ρ = 0.07–0.14).
• Differences in measurement method: Macaque electrophysiology data provided a higher signal-to-noise ratio (SNR) than fMRI, allowing clearer differentiation between untrained baselines and learning rules.

However, it was suggested that differences in the stimulus sets used in the experiments (texture vs. object) may have influenced the change in alignment from V2 to V4.