$ Eddie Berman / Reading (Back)

Some favorites. A combination of things that are relevant to my current work and things that I just think are quite neat :-).

Prior Probabilities. If you don't know which prior to use, what do you do? How can groups acting on the parameters of a distribution inform the prior?

A General Theory of Correct, Incorrect, and Extrinsic Equivariance. When equivariance conditions of the task are not strictly satisfied, do equivariant neural networks fail gracefully? What is the best we can do?

Approximation-Generalization Trade-offs under (Approximate) Group Equivariance. Establishing the limits of equivariant function approximators

To Augment or Not to Augment? Diagnosing Distributional Symmetry Breaking. How the fuck is QM9 canonicalized? Are we giving non equivariant functions an unfair advantage on baselines?

The Lie Derivative for Measuring Learned Equivariance. Vision transformers are more translation equivariant than CNNs. Don't believe me? Then measure!

Deep Learning is Not so Mysterious or Different. title.

Unitary convolution for learning on graphs and groups. Preserve smoothness (as measured by the Rayleigh quotient).