RT @michaelfbonner: Excited to announce that this paper is now out in @ScienceAdvances

RT @michaelfbonner: Can we gain a deep understanding of neural representations through dimensionality reduction? Our new work shows that th….

There’s much more in the paper, including the large influence of universal dimensions on conventional similarity measures like RSA. Check it out!.

The invariance of these representations implies that they are not primarily governed by the details of a DNN’s design but instead by more general principles of natural image representation in vision systems.

This trend was consistently observed across all levels of DNN hierarchies. It was thus not restricted to low-level features in early layers—even higher-level semantic representations appear to be shared across DNNs.

While most DNN dimensions are model-specific, a subset of dimensions consistently emerges across DNNs despite variations in initializations, architectures, and task objectives. These universal dimensions of DNN image representation are also strongly shared with the human brain.

We characterized the universality of latent dimensions in DNN activations by measuring their average decodability across other DNNs. We also computed their decodability from human visual cortex representations measured with fMRI.

Many efforts seek to explain model-brain alignment in terms of the DNN’s architecture and task, but an alternative theory is that brain-like representations reflect universal aspects of natural image representation that emerge in systems with diverse optimization constraints.

Why do varied DNN designs yield equally good models of human vision? Our preprint with @michaelfbonner shows that diverse DNNs represent images with a shared set of latent dimensions, and these shared dimensions turn out to also be the most brain-aligned.

RT @michaelfbonner: New paper out in @PLOSCompBiol. The best deep neural network models of visual cortex do not reduce representations to l….