View on GitHub

Deep Declarative Networks

CVPR 2020 Workshop

Invited Talk: Tractably Modeling with Constraints using Neural ODEs

Ricky TQ Chen
University of Toronto
Neural ordinary differential equations are a class of models that inherently satisfies useful constraints while also being amendable to user-specified ones. They are natural candidates for modeling physical systems, with inductive biases that can be designed, e.g. to satisfy conservation laws. Their invertibility and universality properties make them natural building blocks for general time series and probabilistic models. I will be giving an overview of recent applications of modeling with ODEs, including some of our recent works on tractable probabilistic modeling with normalizing flows and stochastic differential equations.