Reproduce “Brain-like representational straightening of natural movies in robust feedforward neural networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tahereh Toosi, Elias B. Issa
Robustness training on static image-classification tasks can give feedforward neural networks a temporal geometry resembling that observed in biological vision. Adversarially trained and Gaussian-noise-trained ResNets straighten natural movie trajectories, support more accurate feature-space interpolation, and predict macaque V1 responses better than standard networks, although the effect does not consistently extend to higher visual areas.
Key insights
- Adversarial training and random smoothing can induce sustained straightening of natural movie trajectories even without temporal prediction objectives or movie exposure during training.
- Straightening is linked to improved linear invertibility: interpolating between endpoint representations produces images that more closely match actual intermediate movie frames.
- The association between straightening and V1 predictivity is stronger than that between generic robustness and neural predictivity, but robustness alone is insufficient because a biologically inspired robust front end did not show the same effect.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.