Readout Gaussian Processes for Geometric Deep Learning

Lead Research Organisation: Imperial College London
Department Name: Dept of Computing

Abstract

Extending deep learning to work on non-euclidean domains such as graphs and manifolds has been of increasing importance in applications. These include computer graphics computer vision biological networks and computational social science. Moreover it is important to endow such models with a means of quantifying uncertainty while maintaining flexibility and scalability towards large and multiple datasets. In this project we propose Readout Gaussian Processes a probabilistic framework for supervised learning on non-euclidean domains. Flexibility is ensured by means of a deep parametric covariance function, and using stochastic variational inference one is able to perform training and inference at large scale. We conclude by justifying certain methodological choices and mapping out a research program for evaluating these including testing on motivating applications.

Computer graphics, computer vision deep learning

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R512540/1 01/10/2017 30/09/2021
2296938 Studentship EP/R512540/1 01/01/2018 31/12/2021 Eduardo Carvalho