IC-196Models trained on Something-Something-V2 (which contains no faces) show reduced alignment with face-selective FFA compared to the same models trained on Kinetics-400
Christina Sartzetaki, Gemma Roig, Cees G. M. Snoek, Iris Groen
Comparing the same four models trained on Kinetics-400 versus Something-Something-V2, a significant advantage of Kinetics-400 training is observed in FFA and OFA, with a similar pattern in LOC. This effect appears in both shallow and deep layers. Since Something-Something-V2 never shows faces and FFA is the functionally selective region for faces, this demonstrates that a training dataset bias (absence of a stimulus category) transfers to model-brain alignment in the corresponding functionally selective brain area. By contrast, training on Kinetics-710 (an extended version of Kinetics-400) largely has no effect on brain alignment.
Evidence
correlational
Key metric
Significant advantage of Kinetics-400 over Something-Something-V2 in FFA and OFA; Kinetics-710 vs. Kinetics-400 shows largely no effect
Caveat
Only four models are available in both Kinetics-400 and Something-Something-V2 variants, limiting the generalizability of this comparison