IC-413Factors of variation in ImageNet-X are linearly decodable from the second-to-last-layer representations of ImageNet-pretrained ResNet50 and ViT-B/16
Patrik Reizinger, Alice Bizeul, Attila Juhos, Julia E Vogt, Randall Balestriero, Wieland Brendel, David Klindt
The paper tests whether binary factors of variation (defined by human annotators in ImageNet-X) can be recovered by training a linear decoder on the frozen second-to-last-layer features of ResNet50 (d=2048) and ViT-B/16 (d=768), both trained on standard ImageNet classification. In most cases, the linear decoders achieve accuracy significantly above the 50% chance level (Bonferroni-adjusted, κ=0.05/85, 10 seeds of balanced resampling). ResNet50 shows slightly higher decoding performance than ViT-B/16. Some factors (e.g., darker, brighter) are decodable even from randomly initialized models or raw input, while shape information is decodable from the trained models, suggesting shortcut learning may be mitigated after standard training.
Evidence
correlational
Key metric
d= 2048 (ResNet50) vs d= 768 (ViT-B/16); chance level 50%; Bonferroni-adjusted significance κ = 0.05/85; 10 seeds of balanced resampling
Caveat
The factors in ImageNet-X are binary proxies defined by human annotators, not the true latent variables of the data generating process. The authors note this 'only provide[s] some supporting evidence for our theory on real data.' The VMF distributional assumption of the theory may not hold for real-world data.