Turing and FCAI Meetup
Time: 23rd February, 2:30–6:30pm Finnish time, 12.30–4.30pm UK time
Venue: Online (link will be provided to registered participants)
The Alan Turing Institute and the Finnish Center for Artificial Intelligence FCAI signed a memorandum-of-understanding in 2019 and have committed to working together to tackle foundational challenges in machine learning as well as large scale problems in the physical sciences and engineering. There are a lot of synergies between research conducted at the two institutions, and this event will aim to create new connections and collaborations between researchers.
The event will consist of an afternoon of talks highlighting some exciting research developments, as well as smaller sessions in break-out rooms to give an opportunity for researchers to get to know each other. The talks will take place across two sessions: the first session will highlight computational advances in probabilistic machine learning, and the second session will highlight advances in the use of machine learning to tackle problems in the physical and engineering sciences.
The registration is open to anyone affiliated to The Alan Turing Institute or the Finnish Center for AI.
Registration closed (contact organisers for Zoom link).
Schedule
Schedule in Finnish time (UTC+2) and UK time (UTC)
2:30–2:45pm (FIN) | 12.30–12.45pm (UK): Opening remarks
Mark Girolami (Chief Scientist of Turing)
Samuel Kaski (Director of the FCAI and Turing Fellow from the University of Manchester)
Session 1: Probabilistic Machine Learning
2:45–3:25pm (FIN) | 12.45-1.25pm (UK): Chris J. Oates (details below)
3:25pm–4:05pm (FIN) | 1.25pm-2.05pm (UK): Aki Vehtari (details below)
4:05–4:40pm (FIN) | 2.05-2.40pm (UK): Breakout rooms + Break
Session 2: Machine Learning in the Physical and Engineering Sciences
4:40pm–5:20pm (FIN) | 2.40-3.20pm (UK): Laura Ruotsalainen (details below)
5:20–6:00pm (FIN) | 3.20-4.00pm (UK): Pranay Seshadri (details below)
6:00–6:30 pm (FIN) | 4.00-4.30pm (UK): Breakout rooms
Full program
Chris J. Oates: Robust Generalised Bayesian Inference for Intractable Likelihoods
Generalised Bayesian inference updates prior beliefs using a loss function, rather than a likelihood, and can therefore be used to confer robustness when the statistical model is misspecified. This talk considers generalised Bayesian inference with a Stein discrepancy as a loss function, motivated by applications in which the likelihood contains an intractable normalisation constant.
Aki Vehtari: Pareto-k as practical pre-asymptotic diagnostic of Monte Carlo estimates
I discuss the use of the Pareto-k diagnostic as a simple and practical approach for estimating pre-asymptotic reliability of Monte Carlo estimates with use cases in importance sampling, stochastic optimization, and variational inference, which are commonly used methods in probabilistic machine learning.
Laura Ruotsalainen: Deep learning for sustainable smart cities
At present, transport in the European Union contributes to 25% of the total greenhouse gas (GHG) emissions with the majority (70%) of the emissions coming from road traffic (EEA 2018). As most traffic concentrates in urban areas, cities look for practical strategies to make their transport system more efficient and sustainable. Electrification of road transport is the primary technological change needed to meet the carbon reduction targets, but is unlikely to be sufficient. On the other hand, there is a second major technological transformation on-going in road transport—digitalization—which may enable a decrease in CO2 emissions and contribute to the prevention of climate change. Reinforcement learning (RL) is an emerging methodology that is transforming the way many complicated transportation decision-making problems are tackled. However, at present the major deficiency in RL based research addressing autonomous driving and traffic organization is the narrow scope, i.e. focusing on one or two specific aspects of traffic.
In this presentation, we discuss our ongoing research activities aimed at supporting the development of sustainable smart cities through intelligent transport. Firstly, we will present our research optimizing the critical infrastructure required by the intelligent transport and secondly livability of a smart city by traffic organization. Both projects have started by developing feature-based deep Long-Short Term Memory (LSTM) networks for forming input for RL based optimization algorithms. We will present the results from input formation as well as the tuning of a SUMO simulation environment. Thirdly, we will be discussing our recently awarded AIforLEssAuto- project, which will continue the sustainability research by looking at the CO2 emission, and thus climate change effects, of intelligent transport. This project will also be a corner stone of the recently established FCAI virtual laboratory on Smart Mobility and Autonomous Systems, which will address the previously mentioned shortcomings in the RL research.
Pranay Seshadri: Bayesian Assessments of Jet Engine Performance
The operational state of jet engines is typically delineated by scalar (1D) parameter values. These include isentropic and polytropic efficiencies, pressure ratios, and their synthesised scalar pressures and temperatures. The averaging of three-dimensional non-uniform flow-fields to arrive at these stagnation and static values has therefore been the subject of considerable importance. In practice, owing to limited instrumentation and the precision of the sensor apparatus used, there is very little opportunity to adopt thermodynamically sound averaging practices. Instead, more often than not, the area average—computed by weighting each measurement based on the cross-sectional annular area it spans at a given measurement plane—is widely used. However, there is no theoretical basis for this averaging practice, and no guarantee that if one were to add more sensors it would converge to its true value. Additionally, many existing workflows and experimental handbooks on the subject are demonstrably ill-equipped to offer useful advice.
In this talk, I will introduce the Bayesian area and mass averages and detail their computation. These metrics build upon the idea of interpreting a jet engine measurement plane as a Gaussian random field, underscored by a physics-guided statistical methodology that accounts for the radial and circumferential variation seen in engines. Beyond studying the field and its averages at a single measurement plane, the framework can be used to analyse multiple measurement planes via transfer learning yielding an elevated understanding of flow physics. I will also talk about how this Bayesian framework can be deployed for anomaly detection by borrowing ideas from the field of optimal transport. This talk will conclude with a few remarks on the significance of this body of research and its impact on current and future propulsion.
Organisers
1) Ayush Bharti is a postdoctoral fellow at Aalto University and the Finnish Center for AI.
2) Francois-Xavier Briol is a lecturer in Statistical Science at UCL and a Group Leader in Data-Centric Engineering at The Alan Turing Institute.
Thumbnail image by Alan Warburton / © BBC / Better Images of AI / Nature / CC-BY 4.0