Neural Wave Representations

4 October 2023, 2:00 - 2:45 pm

Max Welling

Good neural architectures are rooted in good inductive biases (a.k.a. priors). Equivariance under symmetries is a prime example of a successful physics inspired prior which sometimes dramatically reduces the number of examples needed to learn predictive models. In this work we will try to extend this thinking to more flexible priors in the hidden variables of a neural network. In particular, we will impose wavelike dynamics in hidden variables under transformations of the inputs, which relaxes the stricter notion of equivariance. We find that under certain conditions, wavelike dynamics naturally arises in these hidden representations. We formalize this idea in a VAE-over-time architecture where the hidden dynamics is described by a Fokker-Planck (a.k.a. drift-diffusion) equation. This in turn leads to a new definition of a disentangled hidden representation of input states that can easily be manipulated to undergo transformations. If time allows, I will discuss very preliminary work on how the Schrodinger equation can also be used to move information in the hidden representations. 
Joint work with Andy T. Keller and Yue Song.
To the top of the page