winterschool2021

AI for scientific discovery: an introduction to Accelerate Science

The Accelerate Programme for Scientific Discovery is a new initiative from Cambridge University’s Department of Computer Science and Technology, which will support researchers across the University to use machine learning to advance their research.

What is Machine Intelligence?

Is machine learning really artificial intelligence?

Machine Learning and the Scientific Principle

This talk will focus on how we can formulate beliefs and assumptions mathematically and recover an updated belief from observations. We will reflect on the place machine learning can have in the scientific toolbox.

The Role of Uncertainty in Machine Learning

Humans understand, reason and act by quantifying not only what they know, but also what they do not know. The parallel to this in the machine learning world is quantification of uncertainty, and its propagation across the various components of our machine learning system. This talk will discuss the different sources of uncertainty arising in a modeling scenario, and the tools we can use to capture this uncertainty and use it as part of our machine learning-assisted reasoning.

Hierarchical models for insightful machine learning

Machine learning is a powerful tool to find explanations for data, but not all explanations are created equal. This talk will explore why collaboration with domain experts is critical for the successful application of machine learning in the industrial and scientific domains.

Bayesian Optimisation: Sequential Decision Making Under Uncertainty

In this talk we will focus on global optimisation of black-box functions using Bayesian optimisation, an iterative optimisation technique.

Regression, Causality, Statistical Paradoxes and other Fairy Tales

Regression models can be useful in various ways. In this talk we will focus on how we can use them to compute causal effects, so we can augment our toolkit when reasoning about how the world around us works. We will review some basic concepts of causal reasoning and revisit some common statistical misconceptions that can easily avoided with proper causal thinking.

Inverse Problems in Biology, Deconvolution of Mixed Signals in Spatial Transcriptomics Data, and How to Use Matrix Factorization for Nearly Everything

Inverse problems arise when we want to use data to extract an insight into the inner workings of a system. Such problems are often ill-posed, meaning that there are multiple explanations compatible with the observations and it is therefore necessary to constrain the problem to arrive at a concrete solution.

Universes are Big Data: from geometry, to physics, to ML

We briefly overview how historically string theory led theoretical physics first to algebraic/differential geometry, and then to computational geometry, and now to data science.

A Triangle of Influence: Bringing Together Physics, Pure Mathematics, and Computer Science

Recent advances in machine learning have begun creating new bridges to physics and mathematics that have traditionally existed between the latter two. Given this progress, I will speculate about where we are and where things might be headed, including through the recently launched NSF AI Institute for Artificial Intelligence and Fundamental Interactions.