The RSS Statistical Computing section is hosting a half-day event on Deep Learning at the Royal Statistical Society in London.
This event will include three presentations from academic and industrial researchers actively working on the development and application of deep learning algorithms. The first talk, given by Andreas Damianou (Amazon) will provide an introduction to deep learning, followed by two research-focused talks by Silvia Chiappa (DeepMind) and Ana Namburete (University of Oxford). Titles and abstracts to follow.
1.00 pm -2.00pm: Introduction to Deep Learning (Andreas Damianou, Amazon)
2.00pm - 2.30pm: Annual General Meeting / Coffee/Tea
2.30pm - 3.15pm: Silvia Chiappa (Deep Mind)
3.15pm - 4.00pm: Ana Namburete (University of Oxford)
Andreas Damiano - Introduction to deep learning
In this introduction to deep learning I will firstly discuss the fundamentals of deep neural networks, in particular: motivation, model architecture and optimization. Although deep learning is often described as the optimization of loss functions, in this talk I will offer a complementary view through the lens of probabilistic modelling. It will become evident that the mathematics of deep neural networks are relatively simple; however, the deeper understanding of their operation is only now being generally understood, thanks to the remarkable attention that the community has placed on the field. I will, therefore, discuss some of these recently formed insights: what makes deep neural networks so successful, what are their caveats and what are efficient ways of using them in practice.
Silvia Chiappa - Path-specific counterfactual fairness
The causal framework offers an intuitive and powerful way of reasoning about fairness in machine learning. In this talk, I will discuss the complex scenario in which a sensitive attribute might affect the output of the system along both fair and unfair pathways. I will then
present our work on how to leverage recent advances in deep learning and approximate inference to design a method that can achieve path-specific counterfactual fairness in nonlinear settings.
Ana Namburete - Entering the Cranial Vault: Deep Learning-Based Analysis of the Fetal Brain Ultrasound
Ultrasound (US) imaging is one of the first steps in a continuum of pregnancy care. During the fetal period, the brain undergoes dramatic structural changes, many of which are informative of healthy maturation.The resolution of modern US machines enables us to observe and measure brain structures, as well as detect cerebral abnormalities in fetuses from as early as 14 weeks. Recent breakthroughs in deep learning introduce opportunities to develop bespoke methods to automatically align brain images and then track spatiotemporal patterns of fetal brain development. In this talk, I will summarise my work on the design of appropriate data-driven techniques to extract developmental information from standard clinical US images of the brain.
Registration with is required.