View on GitHub

Machine Learning Tutorials

Imperial College London

Welcome to the homepage of Imperial College London machine learning tutorials. The machine learning tutorials are supported by the Machine Learning Initiative at Imperial College London and the current organizers are Dr. Viktoriia Sharmanska and Dr. K S Sesh Kumar.

The list of past tutorials may be found here.

Schedule 2019-2020

Date Venue Time Title Speaker
20/11/19 Huxley Building, LT 145 14:00 - 16:00 Structured prediction Dr. Carlo Ciliberto
23/10/19 Huxley Building, LT 145 14:00 - 16:00 Advances in machine learning for molecules Dr. José Miguel Hernández-Lobato

Title : Structured prediction

Speaker : Dr. Carlo Ciliberto

Speaker’s Bio : In 2008 Carlo Ciliberto graduated in Mathematics at the University of Roma Tre, Rome, Italy and in 2012 he obtained a PhD in humanoid robotics, computer vision and machine learning at the Italian Institute of Technology, Genova, Italy. He was a Postdoctoral fellow in Poggio Lab at the Massachusetts Institute of Technology from 2012 to 2016 and later a Research Associate at UCL from 2017-2018, where he is now Honorary Lecturer. In 2018 he became a Lecturer at the EEE Department at Imperial College.

Abstract: Modern machine learning offers powerful tools to address supervised problems with linear output spaces (e.g. scalar or vector-valued regression). However, problems with non-linear (and often non-convex or discrete) output spaces are becoming increasingly common. Examples include image segmentation or captioning, speech recognition, manifold regression, trajectory planning, protein folding, prediction of probability distributions or ranking to name a few. These settings are often referred to as “structured prediction” problems, since they require to deal with output spaces that have a complex structure, such as: strings, graphs, images, sequences, manifolds etc. In this tutorial we will introduce the main strategies used to address structured prediction problems. In paritcular we will study: 1) likelihood estimation methods, which are extremely versatile and can be applied to many structured output spaces and loss functions; 2) Surrogate approaches, which need additional care when tailored to a structured problem, but usually lead for tight theoretical characterization of the resulting estimators, namely universal consistency and learning rates. Finally, we will look at recent methods that find a synthesis of these two strategies to achieve both flexibility and soundess of the corresponding algorithms.

Title : Advances in machine learning for molecules

Speaker : Dr. José Miguel Hernández-Lobato

Speaker’s Bio : Since Sep 2016, Hernández-Lobato is a Lecturer in Machine Learning at the Department of Engineering in the University of Cambridge, UK. Before that he was a postdoctoral fellow in the Harvard Intelligent Probabilistic Systems group at the School of Engineering and Applied Sciencies of Harvard University, and a postdoctoral research associate in the Machine Learning Group at the University of Cambridge, UK. He got his PhD from the Computer Science Department in Universidad Autónoma de Madrid (Spain) in 2010. His research revolves around model based machine learning with a focus on probabilistic learning techniques and with a particular interest on Bayesian optimization, matrix factorization methods, copulas, Gaussian processes and sparse linear models.

Abstract: In this tutorial, I will first give a brief introduction to machine learning methods for modelling molecule data, with a strong focus on graph neural networks. After this, I will present some applications of machine learning to molecule data. First, I will focus on the problem of efficiently searching chemical space for new molecules with optimal properties. I will describe how to use recent advances in deep generative models to obtain continuous representations of molecules which allow us to automatically generate novel chemical structures by performing simple operations in a latent space. These methods can then be connected with Bayesian optimization techniques to accelerate the search for new molecules with optimal properties. After this, I will focus on the problem of modelling chemical reactions by predicting electron paths. Chemical reactions can be described as the stepwise redistribution of electrons in molecules. As such, reactions are often depicted using “arrow-pushing” diagrams which show this movement as a sequence of arrows. I will describe an electron path prediction model to learn these sequences directly from data and show that the model recovers a basic knowledge of chemistry without being explicitly trained to do so. Finally, I will describe a generative model for molecules that gives a synthetic route to the molecules generated.