Validating Learning Rules Using Human and Macaque Brain Data

This is a study that applied and compared 5 learning rules—BP, PC, STDP, FA, and others—to human fMRI data and macaque electrophysiology data.

The key findings are as follows:

• V1/V2 Regions: STDP (ρ ≈ 0.30) and PC (ρ ≈ 0.28) showed high alignment in the macaque V1/V2 regions, consistent with the human V1 results. This confirmed that the alignment pattern in early visual regions is qualitatively preserved across species.
• IT Region: Alignment tended to be determined more by the model's capacity than by a specific learning rule. Pretrained ResNet-50 recorded an alignment of ρ ≈ 0.25 in the macaque IT region.
• Data Precision: Macaque electrophysiology data provided a higher signal-to-noise ratio (SNR) than fMRI, clearly revealing differences between learning rules that were difficult to distinguish with fMRI.

As a limitation of the study, it was noted that the drop in alignment when moving from V2 to V4 may have been influenced by differences in the stimulus sets used (textures vs. objects).