May 9th Lightning Speakers
Decomposition of weighted directed networks into acyclic plus circulating by Professor Robert MacKay
About: The main motivation is payment networks, where if A owes B owes C owes A then in principle they could each write off the minimum of the amounts owed, thereby reducing their liquidity requirements and transfer charges. One could reduce the payment network to an acyclic one.
The question is what is the space of possible acyclic results and what might be a sensible objective function to optimise over it. This is work in progress with Bazil Sansom (WBS), Yijie Zhou (MathSys CDT) and Marc Homes Dones (Maths postgraduate).
Inegrating Decision Support for Coupling together Massive Models by Professor James Smith
About: Increasingly often massive dynamic probabilistic forecasting models need to be coupled together to provide coherent outputs to support decisions. James will outline how, over the last two years, he has led teams that have harnessed recent formal methodologies through two large applications - work resourced through the Turing - Warwick collaboration.
The first decision support methodology is designed to help local government to decide between different competing portfolios of carbon zero heating options. The second is designed to help three different asset owners to co-ordinate their strategies to mitigate of climate change induced increased flooding risk through developing a digital twin that could perform uncertainty handling.
Detecting and localising changes in different environments by Dr Yi Yu
About: In this talk, Dr Yi Yu will give a brief overview of the change point analysis area, including some motivating examples and the fundamental limits which their group are interested in.
Find out more about Yi's research.Link opens in a new window
Statistical and mathematical methods in personalised medicine by Dr Deepak Parashar
About: In this talk, Dr Parashar will present his current research on statistical methodology and mathematical techniques underpinning personalised medicine, where the right patient gets the right treatment at the right time.
During his talk, he will highlight recent developments in novel designs of biomarker-driven clinical trials and emerging ideas on geometric visualisation of multidimensional health data, and how these align with Alan Turing Institute’s priority areas.
Machine Reasoning for Language Understanding by Professor Yulan He
About: Enabling machines to understand human languages requires equipping AI systems with reasoning capabilities. Reasoning is not necessarily achieved by making logical inferences.
In our context, it refers to being able to manipulate previously acquired knowledge in order to enhance the understanding of new information encountered. In her talk, she will outline our recent progress in machine reasoning, including incorporating common sense knowledge for emotion detection in dialogues, event semantic reasoning for answering questions relating to real-world events reported in news, and evidence reasoning for claim veracity assessment. Professor He will conclude her talk with challenges and potential future research directions.
Rethinking (Deep) Learning on Devices: From Low-Power Sensing to Zero-Cost Automated ML by Dr Hongkai Wen
About: The unprecedented wealth of data has led to new computation paradigms based on deep learning. However, to make these computations possible, especially on devices, we need to rethink the design of both our algorithms and systems, and tailor them to keep up with the ever-increasing computation demand.
In this talk, he will give a brief overview of several our projects related to on-device AI, from tiny perceptual super-resolution, to ultra low power event based motion sensing and zero-cost AutoML.
Nash Neural Networks by Professor Matthew Turner
About: We propose Nash Neural Networks (NNN) as a new type of Physics-informed Neural Network that is able to infer an underlying utility from observations of rational individuals in a differential game with a Nash equilibrium. We assume that the dynamics for both the population and the individual are known, but not the payoff function. We construct our network in such a way that the Euler-Lagrange equations of the corresponding optimal control problem are satisfied and the optimal control is self-consistently determined. In this way, we are able to learn the unknown payoff. function in an unsupervised manner. We have applied NNN to study the optimal behaviour during epidemics, in which individuals can choose to socially distance depending on the state of the pandemic and the cost of being infected. Training our network against synthetic data for a simple SIR model, we showed that it is possible to reproduce the hidden payoff function, in such a way that the game dynamics are respected. Our approach will have far-reaching applications, as it allows one to infer utilities from behavioural data, and can thus be applied to study a wide array of problems in science, engineering, economics and government planning.