Gatsby Computational Neuroscience Unit CSML University College London

Gatsby Computational Neuroscience Unit
Sainsbury Wellcome Centre
25 Howland Street
London W1T 4JG UK

+44 (0)7795 291 705

bottom corner


Arthur Gretton I am a Professor with the Gatsby Computational Neuroscience Unit, part of the Centre for Computational Statistics and Machine Learning at UCL. A short biography.

My research focus is on using kernel methods to reveal properties and relations in data. A first application is in measuring distances between probability distributions. These distances can be used to determine strength of dependence, for example in measuring how strongly two bodies of text in different languages are related; testing for similarities in two datasets, which can be used in attribute matching for databases (that is, automatically finding which fields of two databases correspond); and testing for conditional dependence, which is useful in detecting redundant variables that carry no additional predictive information, given the variables already observed. I am also working on applications of kernel methods to inference in graphical models, where the relations between variables are learned directly from training data.

Recent news

Machine Learning Summer School to take place July 2019 in London, co-organised with Marc Deisenroth.
Learning deep kernels for exponential family densities: a scheme for learning a kernel parameterized by a deep network, which can find complex location-dependent local features of the data geometry. Code, talk slides, and high level explanation.
On gradient regularizers for MMD GANs, NIPS 2018. A new gradient regulariser for MMD GANs with state-of-the-art performance (as of June 2018) on 160x160 CelebA and 64x64 Imagenet. Code and talk slides.
Informative Features for Model Comparison, NIPS 2018. When comparing complex generative models in high dimensions, the question to ask is not "which model is correct" (neither), or "which model is better," but rather "where does each model do better than the other?" Code
BRUNO: A Deep Recurrent Model for Exchangeable Data, NIPS 2018. A deep generative model which is provably exchangeable, meaning that the joint distribution over observations is invariant under permutation. The model does not require variational approximations to train. Used for generalisation from short observed sequences. Code
• Course slides and videos for the 2018 Machine Learning Summer School Madrid, and the 2018 Data Science Summer School Paris, now linked in the teaching page.
Talk slides from the ICML 2018 Workshop on Theoretical Foundations and Applications of Deep Generative Models. Focus is on the papers On gradient regularizers for MMD GANs and Demystifying MMD GANs.

Older news

Demystifying MMD GANs, ICLR 2018. Wasserstein GANs, MMD GANs, and Cramer GANs all have the exact same bias properties: all gradients are unbiased, but the critic may still give biased losses when trained. Also: simpler discriminator networks compared with WGANs, dynamic adaptive learning rate adjustment; a new Kernel Inception Distance. Code
Conditional infinite exponential family, AISTATS 2018. learns conditional density models which can be sampled by HMC. Talk slides and code
Efficient density estimator, infinite dimensional exponential family Oral presentation, AISTATS 2018. Contains a comparison with a score estimator based on autoencoders. Talk slides and code
NIPS 2017 best paper award: a linear time kernel goodness-of-fit test . Gives a linear time test for assessing the quality of a model, compared with a reference sample. Code, including a demonstration notebook from the ML Train NIPS workshop

bottom corner