conftrace_

2019 NeurIPS NeurIPS 2019

Augmented Neural ODEs

Abstract

We show that Neural Ordinary Differential Equations (ODEs) learn representations that preserve the topology of the input space and prove that this implies the existence of functions Neural ODEs cannot represent. To address these limitations, we introduce Augmented Neural ODEs which, in addition to being more expressive models, are empirically more stable, generalize better and have a lower computational cost than Neural ODEs.

📈 Trend Setter - Neural Networks

🧭 Keyword Pioneer - neural ordinary differential equation

🐣 Hot Topic Early Bird - neural ordinary differential equation

Authors

Emilien Dupont, Arnaud Doucet, Yee Whye Teh

Topics

Machine Learning > Core Methods > Representation Learning Machine Learning > Optimization & Theory > Neural Network Optimization Deep Learning > Architectures > Neural Networks Deep Learning > Learning Types > Representation Learning Machine Learning > Core Methods > Neural Networks

Keywords

representation learning deep learning neural ordinary differential equation topology preservation model expressivity augmented neural ode

More from NeurIPS 2019

All NeurIPS 2019 papers →

Generalization in multitask deep neural classifiers: a statistical physics approach

Causal Regularization

Locality-Sensitive Hashing for f-Divergences: Mutual Information Loss and Beyond

Efficient Smooth Non-Convex Stochastic Compositional Optimization via Stochastic Recursive Gradient Descent

Regularizing Trajectory Optimization with Denoising Autoencoders

Multi-Criteria Dimensionality Reduction with Applications to Fairness