Robust feedforward CNNs straighten natural-movie representational trajectories

- Representational curvature (mean angle between vectors connecting representations of nearby movie frames, as in Hénaff et al. 2019) of natural-movie sequences is lower, in layers of ResNet-50, for adversarially trained (L2 and Linf) and noise-trained (random-smoothing base classifier) ImageNet models than for a standard-trained ResNet-50 and other non-robust ResNet-50 baselines (Fig. 2). [toosi-issa-2023-representational-straightening-robust-cnns] - Linear interpolation between the representations of the first and last frame of a movie yields in-between frames more similar to the true frames (higher SSIM) for adversarially trained and random-smoothing ResNet-50s than for a standard ResNet-50 (Fig. 3). [toosi-issa-2023-representational-straightening-robust-cnns] - The straightening arises without a predictive objective or training on movies: the authors state that "it is not necessary to utilize predictive objectives or train directly on natural movie statistics" to obtain it, and layers with more straightening better predict macaque V1 responses (Fig. 4); the measurements are observational (the authors compare existing training regimes rather than intervening on the geometry). [toosi-issa-2023-representational-straightening-robust-cnns]