Talks
These are some of the talks I’ve given in the past:
Gaussian processes for non-Gaussian likelihoods
Gaussian Process Summer School 2022
Following the online format in 2021, I was invited back to lecture at the 2022 summer school, this time finally back in person. Having a live audience is a lot more engaging, I’m very grateful for for the interaction. Giving a talk multiple times is a great opportunity for refining in different ways how to best get the content across!
[ Slides ]
Gaussian processes for non-Gaussian likelihoods
Lifting Inference with Kernel Embeddings (LIKE22) winter school
Another take on the tutorial I previously presented at the Gaussian Process Summer School.
Encoding prior knowledge through Gaussian processes
Methods of Machine Learning, Universität Tübingen, 2021
Abstract: With lots of data that you don’t know much about, just fitting a reasonably flexible model might be good enough. But what if you don’t have as much data, or you already have some knowledge about its properties? Encoding prior knowledge makes a model intentionally less flexible, and allows it to generalise better from less data. In this talk I will give a short introduction to how we can encode prior knowledge in Gaussian processes, for example invariances in the input space.
[ Slides ]
Gaussian processes for non-Gaussian likelihoods
Gaussian Process Summer School 2021
Abstract: Functions that we model with Gaussian processes have, by construction, Gaussian marginals. However, real world data is often better described by other distributions: whether this is classification (Bernoulli or Multinomial distributions), observations with some constraints such as positivity (Gamma or log-Normal distributions), or discrete observations such as counts (Poisson or Negative-Binomial distributions). In this lecture, I show how we can combine a latent function (which we will describe using a Gaussian process) and non-Gaussian likelihoods. I give an introduction to various different ways of approximating the (non-Gaussian, and therefore not analytically computable) posterior: Laplace approximation, Expectation Propagation, Variational Inference (Variational Bayes), and Markov Chain Monte Carlo (MCMC).
[ Slides ]
The ICML Debate: Should AI Research and Development Be Controlled by a Regulatory Body or Government Oversight?
ICML 2021
Not exactly a talk: at ICML 2021 I argued with (and against) Charles Isbell, Michael Kearns, Rich Sutton, Steve Roberts, Suchi Saria, Shakir Mohamed, and Martha White in “British Parliament Style”. The first time I had done anything like this, it was stressful and I can think of so many ways to do better … but it was great to be part of it nonetheless!
[ Video ]
Gaussian processes for fun and profit: Probabilistic machine learning in industry
Foundational AI Seminar Series, University College London Centre for Artificial Intelligence, 2020
Abstract: When companies, whether start-ups or big corporations, talk about “machine learning” they usually mean some kind of neural network model. Not always though: I will talk about why instead we put a lot of our efforts on probabilistic models built using Gaussian processes. When a Machine Learning course briefly covers Gaussian processes, you might go away thinking they’re just basis function interpolation, only apply when the noise is Gaussian, and don’t scale to larger datasets. Here I will discuss why these are misconceptions and show why Gaussian processes are both interesting and useful in practical applications.
[ Slides | Interactive visualisation ]
Multiple dispatch in the inducing variable and multi-output framework in GPflow
Deep Structures workshop at the Finnish Center for Artificial Intelligence, 2019
This was a talk about the importance of good software abstractions in writing composable, re-usable research code, and how we make use of multiple dispatch to achieve this in GPflow.
[ Slides ]
Invariances in Gaussian processes and how to learn them
Structurally Constrained Gaussian Processes workshop at GPSS 2019
Abstract: When learning mappings from data, knowledge about what modifications to the input leave the output unchanged can strongly improve generalisation. Exploiting these invariances is commonplace in many machine learning models, under the guise of convolutional structure or data augmentation. Choosing which invariances to use, however, is still done with humans in the loop, through trial-and-error and crossvalidation. In this talk, we will discuss how Gaussian processes can be constrained to exhibit invariances, and how this is useful for various applications. We will also show how invariances can be learned with backpropagation using tools from Bayesian model selection.
Jointly with Mark van der Wilk.
[ Slides ]
Bayesian modelling of large-scale spatiotemporal discrete events: Gaussian processes, Cox processes and Fourier features
Infectious Disease Epidemiology Seminar Series, Imperial College London, 2018
Abstract: Bayesian inference allows us to build probabilistic models in which we derive not just point estimates, but also uncertainty quantification from our observed data. An underrepresented area in this field of research is the modelling of point processes: discrete events such as disease cases in epidemiology, locations of trees in ecology or distribution of taxi pickup requests in smart cites. I will talk about how we can use Gaussian processes to infer the rate function that describes the intensity of events across the spatiotemporal domain in a Cox process. I will give an overview over different approaches and their advantages and disadvantages, including recent work on using Fourier features (a representation of the function in the spectral domain) to scale up point process inference to larger data sets.
[ Slides ]
Many-body coarse grained interactions using Gaussian approximation potentials.
Advances in Data Science Seminar, University of Manchester, 2018
Abstract: We introduce a computational framework that is able to describe general many-body coarse-grained (CG) interactions of molecules and use it to model the free energy surface of molecular liquids as a cluster expansion in terms of monomer, dimer, and trimer terms. The contributions to the free energy due to these terms are inferred from all-atom molecular dynamics (MD) data using Gaussian Approximation Potentials, a type of machine-learning model that employs Gaussian process regression. The resulting CG model is much more accurate than those possible using pair potentials. Though slower than the latter, our model can still be faster than all-atom simulations for solvent-free CG models commonly used in biomolecular simulations.
[ Slides ]