SPAAM Seminar Series 2022/23
For information on current updated seminar talks, visit the Statistics, Probability, Analysis and Applied Mathematics (SPAAM) seminar series websiteLink opens in a new windowLink opens in a new windowLink opens in a new window.
The Statistics, Probability, Analysis and Applied Mathematics (SPAAM) seminar series will take place on Thursdays between 3-4pm in room B3.02 and virtually on the SPAAM Microsoft Teams ChannelLink opens in a new windowLink opens in a new window. It will host a variety of talks from PhD students involved in applied mathematics research at Warwick and invited guests from other institutions (see the bottom of this page for the talk abstracts!).
Each seminar will usually host two speakers (unless otherwise stated) with each talk taking around 15-20 minutes with 5-10 minutes of questions afterwards. Speakers and committee members will hang around for some time after the talks for social tea/coffee and further questions. Please do contact one of the committee if you would like to join and be added to the MS Teams channel. Note that these talks may be recorded for later viewing on our Youtube channelLink opens in a new windowLink opens in a new window so do join with audio and video off if you don't wish to feature!
If you would like to give a talk this term, please contact Jack Buckingham (email@example.com) or Olayinka Ajayi (Olayinka.Ajayi@warwick.ac.ukLink opens in a new windowLink opens in a new window) and we will find you a slot!
Term 3 Talks
|4th May 2023 (Week 1)||/|
|11th May 2023 (Week 2)||Alex KayeLink opens in a new window & Phoebe AsplinLink opens in a new window
Link opens in a new windowLink opens in a new window
|18th May 2023 (Week 3)||Yi Ting LooLink opens in a new window & Mozhdeh ErfanianLink opens in a new window|
|25th May 2023 (Week 4)||Zak Ogi-GittinsLink opens in a new window & Phurinut SrisawadLink opens in a new window|
|2nd June 2023 (Week 5)||Jack O'ConnorLink opens in a new window|
|9th June 2023 (Week 6)|
|16th June 2023 (Week 7)|
|23th June 2023 (Week 8)|
|30th June 2023 (Week 9)|
|7th July 2023 (Week 10)|
Week 5 Jack O'Connor - Incrementally verifiable computation - verifying the meaning of life, the universe, and everything
Say humanity is carrying out some very important computation that will require so much time to run that it will span multiple generations of people. How can future generations be convinced that the computation is still running correctly, so that they know maintaining it isn't a waste of their valuable resources?
In this talk I will introduce you to the notion of incrementally verifiable computation (IVC), which allows us to verify the correctness of long-running computations. IVC has a myriad of applications (apart from multi-generational computations) which include image authentication, blockchain and verifiable delay functions. During the talk I will touch on a number of areas including cryptographic proof systems, applications of IVC, how to construct IVC, some very interesting questions about polynomials and, of course, how to verify the meaning of life, the universe, and everything.
Week 4 (talk 1) Zak Gittins - Estimating Reproduction Numbers (No SIR models allowed!)
Reproduction numbers (Rt) are important statistics in multiple disciplines (e.g Population growth, Information dissemination, etc.) but perhaps most commonly in Epidemiology. Inferring these statistics is challenging given limited data, particularly when the incidence data is resolved on time-intervals that are long compared to the pathogen’s transmission time. In this talk, I will give an overview of reproduction numbers, a detailed illustration of the most popular inference method used in the literature, and some progress I have made in my PhD regarding inferring Rt with problematic data.
Week 4 (talk 2) Nut Srisawad - Tutorial on identifying the best among possible options
The problem of identifying the best option from stochastic samples has been studied broadly for many years. It can be applied in various fields, for example A/B testing to select the best product for customers, clinical trials to investigate the best treatment for patients, hyperparameter tuning for machine learning, etc. To identify the best solution efficiently, allocation policies are designed to iteratively sample the alternative that promises to provide the most valuable information, based on statistical analysis. In this talk, I will introduce some efficient allocation policies from the literature to decide the next sampling. Then, I will present the current work when the environment of the samples can change.
Week 3 (talk 1) Yi Ting Loo - Reaction-diffusion models of neuruloid pattern formation
During tissue morphogenesis and organ development, cell fates appear robustly self-organised in space and time. However, the underlying principle of this process remains poorly understood, particularly the role of boundary constraints. This project utilises neuruloid, a human embryonic stem cell culture system that recapitulates the early stages of the human nervous system development, as a model system to study this process. Initial imaging results show a striking difference in cell morphology when the boundary constraints of these tissues are varied, showing evidence of geometrical effects. In this talk, we will discuss how reaction-diffusion systems can be utilised to develop phenomenological models that can reproduce neuruloid pattern formation seen in imaging data. I will also give a brief overview of the background of my PhD project, the main motivations, and future work.
Week 3 (talk 2) Mozhdeh Erfanian - Asymptotically Modelling of Plastic Deformation during Cold Rolling of Sheet Metal
Cold rolling is a metal forming process where a strip of metal passes between two rollers and comes out thinner. Understanding the stress and strain in the material resulting from the roll pressure is essential not only for process design but also for material design. Here, we present a multiple-scale asymptotic model for cold rolling under the assumption of large rollers and a thin sheet. This model can successfully predict the stress and strain distribution, including through-thickness stress oscillations, and is shown to agree well with Finite Element simulations.
Week 2 (talk 1) Phoebe Asplin - Vector-borne disease and climate change: forecasting Aedes aegypti mosquito populations
The mosquito that spreads dengue is quite picky about where it lives. The environmental conditions must be neither too hot nor too cold and the amount of rainfall has to be just right. We describe the ecological niche of our Goldilocks like mosquito, Aedes aegypti, and use this information to characterise environmental suitability for the mosquito up to the year 2100. Heavy emphasis will be placed on the unpredictable nature of our climate and how this will cause large uncertainties in the future. Time permitting we talk about how to use these mosquito forecasts to predict the basic reproduction number of dengue and which areas are likely to be most at risk moving forwards. and the risk of dengue
Week 2 (talk 2) Alex Kaye - Investigating the effect of symptom propagation on health economic outcomes
Symptom propagation is where an individual's symptom severity depends, at least partially, on the symptom severity of the individual that infected them. There is growing evidence that this is a phenomenon that occurs for many respiratory pathogens, including SARS-CoV-2 and Influenza. In this talk, I will discuss the pathways through which we think symptom propagation occurs and what impact it has on epidemiology outcomes. I will also explore how symptom propagation impacts the effectiveness of intervention strategies and the challenges that come with separating symptom propagation from other factors that affect cost-effectiveness.
Term 2 Talks
|12th January 2023 (Week 1)||Yueting HanLink opens in a new window|
|19th January 2023 (Week 2)||Minghan ZhangLink opens in a new window & Jummy McKendrickLink opens in a new window
Link opens in a new windowLink opens in a new window
|26th January 2023 (Week 3)||Abi ColemanLink opens in a new window|
|2nd February 2023 (Week 4)||social|
|9th February 2023 (Week 5)||Yijie ZhouLink opens in a new window|
|16th February 2023 (Week 6)||Jonna RodenLink opens in a new window & Jake ThomasLink opens in a new window|
|23rd February 2023 (Week 7)||social|
|2nd March 2023 (Week 8)||Dr Radu CimpeanuLink opens in a new window|
|9th March 2023 (Week 9)||Francesca BasiniLink opens in a new window & Richard FoxLink opens in a new window|
|16th March 2023 (Week 10)||George Watkins Link opens in a new window|
Week 10 Generating a Graph Colouring Heuristic with Deep Q-Learning and Graph Neural Networks - George Watkins
The graph colouring problem consists of assigning labels, or colours, to the vertices of a graph such that no two adjacent vertices share the same colour. In this talk I will present my algorithm, ReLCol, which uses Reinforcement Learning, together with a graph neural network for feature extraction, to learn a heuristic for the graph colouring problem.
Week 9 (talk 1) Likelihood-Free Inference approaches for the estimation of catastrophe models with applications in developmental biology - Francesca Basini
Cell differentiation, the process by which immature cells develop into specialised ones, is influenced by their gene expression profiles. As genes activate and deactivate, they determine the cell's characteristics and type. Understanding these complex decision-making mechanisms is difficult and the nature of the process is high-dimensional. However, data often lie on or near lower-dimensional manifolds in the feature space.
The quasi-potential landscape provides a simple and intuitive framework for modeling cell development. By creating archetypal models controlled by a few parameters, we can represent the differentiation phenomenon in a lower space. These models also have an appealing connection to the Waddington's epigenetic landscape metaphor.
In this talk, we introduce a few simple models built using catastrophe theory and focus on estimating their parameters. We developed two estimation approaches in the context of likelihood-free inference. The first approach extends the Linear Noise Approximation method, approximating the likelihood with a lower-space projection and using a Population Monte Carlo algorithm for inference. The second approach uses Approximate Bayesian Computation and considers two suitable discrepancy measures between simulated and observed data: the exact Wasserstein distance and the Energy Score. We present the results of both estimation procedures on synthetic data scenarios and discuss the strengths and weaknesses of each approach. Finally, we briefly suggest further ideas that are currently under development.
Week 9 (talk 2) Pedestrian and autonomousvehicle interaction - Richard Fox
Cleanly incorporating behavioural science when assimilating human feedback from autonomous vehicle interaction in computational reinforcement learning.
Week 8 Talking the talk: dos, don'ts, and a few ideas in between - Dr Radu Cimpeanu (Guest Speaker)
In this (interactive!) session we will try to find the recipe for a memorable scientific talk together. Putting different hats on, from audience member to presenter to organiser/chair, we will try to get through the Why? How? and What next? of crafting scientific presentations. There is no one size fits all approach to this process, but through examples and practical points we will delve into both content and delivery aspects, and will hopefully leave the room excited to implement some new ideas into our next one.
Week 6 (talk 1) Computational Modelling and Optimal Control for Interacting Particle Systems - Jonna Roden
What do beer brewing, bird flocking, printing, and nano-filtration have in common? Find out at the SPAAM Seminar this week! There I will explain how to build a mathematical model that describes all the things I mentioned above, and more! Once we have this model (and found a way to solve it), we will ask further questions: How do we get the yeast to sediment quicker, so the brewing process is sped up? How do we get the ink to dry uniformly on the paper when we print? Or put in a broader context: What is the optimal problem setup (the thing I can control) that will get me the closest to my desired outcome?
In my research I am working on exactly these questions and if you’re interested in the official description: I am working on numerical solutions to integro-PDEs, and optimal control of interacting particle systems. But don’t worry; you can definitely come along and learn something new without knowing anything (yet!) about my field!
Week 6 (talk 2) Less Labelling: Active Learning with Possibility Theory - Jake Thomas
Active Learning is a form of machine learning in which the agent must actively query the user for labels for given data points. Typically, the agent has access to a large set of unlabelled examples, and due to the expensive labeling process, it must learn from as few labels as possible. Effective active learning requires an accurate model of uncertainty across the input space. In this talk, I will discuss how this can be achieved with possibility theory.
I will briefly introduce the fundamentals of possibility theory before developing the theory of possibilistic Gaussian processes. Finally, I will show how possibilistic GPs can be intuitively used to perform active learning and demonstrate the promising performance we have achieved on baseline problems.
Week 5 Compression and systemic risk: From the perspective of trophic analysis - Yijie Zhou
After the 2007–2008 financial crisis, regulators make different policies and look for ways to reduce the risks within the derivative and global financial market. Portfolio compression is one of the approaches used. It is a post-trade tool that removes redundancy in OTC derivatives networks. This relates to the reduction of trophic incoherence, a concept from trophic analysis, of the pre-compression networks. In this talk, I will briefly discuss how trophic analysis can be used to study compression and risk associated with it.
Week 3 The Spread of Disease in Bees - Abi Coleman
Between climate change, pesticide use and developing diseases, managed honeybees are facing increasing pressures. Varroa destructor is a mite that spreads diseases between honeybees, and has spread across the world since it switched hosts around 1970. However, its spread can be used as a valuable case study for the spread of a bee disease. In this talk I examine a method for modelling the spread of varroa, discuss the issues with this and how they might be mitigated.
Week 2 (talk 1) Facial Expression Analysis on Pain and Empathy for Pain Detection - Minghan Zhang
Sensory and emotional experiences such as pain and empathy are relevant to mental and physical health. The current drive for automated pain recognition is motivated by a growing number of healthcare requirements and demands for social interaction make it increasingly essential. Faces are a special class of stimuli in that they are one of the most expressive perceptual signals for the emotional state of another person. Indeed, facial expressions play a central role in social interactions and can elicit rapid responses in the observer. Especially, observing someone expressing emotion or someone in pain tends to elicit empathy.
We would like to investigate: How do people's facial expressions when they are in pain or empathising with someone in pain differ or resemble one another? Does the intensity of empathy for pain shown on facial expressions differs when observers taking different perspectives, i.e., self-perspective and others-perspective? Does the intensity of empathy for pain shown on facial expressions differ with respect to familiarity with the person observed in pain? Through our research, we can explore affective AI by both Psychology and Computer Science which may help to unravel how humans understand others based on sensory and emotional states. Beyond that, artificial pain empathy could enable an AI assistant to become more socially acceptable and generate positive interpersonal interaction.
Week 2 (talk 2) Modelling seasonality of Lassa fever cases in Nigeria - Jimmy McKendrick
Lassa fever (Lf) is a viral haemorrhagic disease endemic to West Africa and is caused by the Lassa mammarenavirus (LASV). The rodent Mastomys natalensis serves as the primary reservoir and its ecology and behaviour have been linked to the distinct spatial and temporal patterns in incidence. Nigeria experienced an unprecedented epidemic which lasted from January until April of 2018 and has been followed by subsequent epidemics of Lassa fever in the same period every year since. While there have been previous works modelling the case seasonality within Nigeria we have not seen a model capturing both the seasonal variation in the reproduction of the zoonotic reservoir and its effect on case numbers while being fitted with an approximate Bayesian computation scheme to the case data from 2018--2020 supplied by the NCDC.
In this paper we use a periodically forced seasonal nonautonomous system of ordinary differential equations as a vector model to demonstrate that the population dynamics of the rodent reservoir may be responsible for the spikes in observed cases in humans. The results support that attention and efforts should be concentrated at the end of the year to effectively combat Lassa fever in Nigeria.
Week 1 A Novel Understanding of Vaccination Views Competition - Yueting Han
Social media platforms play a significant role in vaccination views polarization, drawing the public's attention especially after the COVID-19 pandemic. Recent literature shows that the pro-vaccine group is losing this battle on Facebook in providing less guidance to the neutral group. Using the same real-world temporal dataset, our research instead observes that, under some data-driven modelling assumptions on the growth rate of pro- and anti-vaccine pages followers, while the anti-group is estimated to have the largest number of followers for the next few decades (in line with prior research), the pro-vaccine later exceeds its number of followers in our projection. To explain such novel prediction results, we investigate the information diffusion patterns of each vaccine group by studying a mesoscale structure (lying between the microscale and macroscale) called the "bow-tie structure" of the directed online social networks.
|20th October 2022 (Week 3)||Juan UngreddaLink opens in a new windowLink opens in a new window & Paul Kent Link opens in a new windowLink opens in a new window|
|27th October 2022 (Week 4)||James PriceLink opens in a new windowLink opens in a new window & Andrew NugentLink opens in a new window|
|3rd November 2022 (Week 5)||Social Event (Common Room)|
|10th November 2022 (Week 6)||Yiming MaLink opens in a new windowLink opens in a new window|
|17th November 2022 (Week 7)||Jack BaraLink opens in a new window & Rachel SeibelLink opens in a new window|
|24th November 2022 (Week 8)||Adam SmithLink opens in a new window|
|1st December 2022 (Week 9)||Social|
|8th December 2022 (Week 10)||Olayinka AjayiLink opens in a new window|
Week 10 Deep Spatio-temporal Models for Surveillance - Olayinka Ajayi
Have you ever wondered how surveillance is achieved by most governments of the world? How they are able to regularly and accurately monitor your physical activities? Though conversations on politics, policies and ethical correctness are relevant for this discussion, rather we would focus on the design of the machine learning models that helps achieve surveillance. In particular, I would be discussing my research on human action recognition using graph neural networks in addition to temporal models.
Week 8 Understanding Neuronal Dynamics behind Fluorescent Calcium Imaging - Adam Smith
Recent advances in fluorescence imaging permit large-scale recording of neural activity. Unlike other methods, fluorescent imaging captures the detail of local neuronal dynamics at mesoscale. Whilst an attractive recording method, this imaging requires pixel-by-pixel resolution that then only capture the combination of fluorescent signals emitted from many neurons. As a consequence, we lose information at single cell level that may hold key information in understanding complex neuronal dynamics.
In this talk, I outline a model that uses our knowledge of calcium ion transport in neurons and how this relates to fluorescent recordings. I’ll then explain how to extend this framework such that it more closely resembles the signals in our imaging data and discuss how this model can help us understand the neuronal dynamics in experiments of cortical spreading depression and seizures.
Week 7 (Talk 1) - Urban migration promotes cooperation in spatial social dilemmas - Jack Bara
In this talk we discuss spatial public goods games in which agents either pollute (defectors) or clean (cooperators) their local area and can migrate to empty sites within range. We ask whether migration alone reduces the pollution felt by individuals, even keeping the number of polluters constant. Analytically and through agent-based simulations, we show that polluters encourage eco-friendly neighbours to migrate away, eventually clustering with other cooperators. We conclude that migration ultimately reduces the pollution felt per-capita, while providing an extra advantage for cooperation. Our results reveal that movement alone can have a positive impact on long-term payoffs in those multiagent systems where space is a key feature of the interaction.
Week 7 (Talk 2) - Incorporation of human behaviour in infectious disease models which include intervention strategies - Rachel Seibel
Human behaviour is an integral driver in emerging infectious disease systems – from humans, animals, to plants. This talk will serve as a PhD research proposal, where we will explore a roadmap for three years of research in the field of behavioural epidemiology. In human disease systems, we ask the question: In the context of behavioural-epidemiological modelling of human infectious disease dynamics, how does individual-level information about the state of the outbreak affect epidemic impact? In livestock and crop systems, we are interested in how farmer behaviours regarding intervention strategies may impact epidemic severity. We consider the challenge of determining which human behaviours should be included in mathematical models of pathogen dynamics, and we propose a pathway to develop an interdisciplinary framework for future behavioural models.
Week 6 - Multi-view crowd counting - Yiming Ma
There have been great advances in single-view crowd counting algorithms. Nevertheless, state-of-the-art methods still cannot handle occlusion at a satisfactory level. Thus, multi-view crowd counting, in which for every fixed scene there are multiple cameras deployed at different views, has been proposed, but it also consequently introduce another tricky problem – how to effectively fuse the information from different perspectives.
Week 4 (Talk 1) - Exploring the role of the potential surface in the behaviour of early warning signals - Andrew Nugent
The theory of critical slowing down states that a system displays increasing relaxation times as it approaches a critical transition. These changes can be seen in statistics generated from timeseries data, which can be used as early warning signals of a transition. Unfortunately, the observed behaviour of early warning signals often differs from what is expected. This talk will first discuss several potential causes of these differences, including the alternative theory of critical speeding up. We then present a characterisation of the behaviour of early warning signals in terms of several key features of a system: namely a system’s potential surface and noise around a quasi-steady state. Finally, we will discuss an ‘equation-free’ method to obtain these features from timeseries data and give a demonstration on an adapted SIS model. We show that combining the ‘equation-free’ method and characterisation of early warning signals provides a new route to explaining their complex behaviour.
Week 4 (Talk 2) - Decision Making for Repeated Games - James Price
In situations where an agent has to make decisions with random outcomes there is no definitive metric for accurately evaluating different policies. For example, when considering whether to put money into a bank account or invest in the stock market it is impossible to know which option will result in the greatest wealth gain by some pre-determined time in the future. For an infinite time-horizon we can use the law of large numbers to calculate a metric known as the growth rate. The growth rate is a real number which represents the average wealth gain and therefore allows us to definitely rank random processes. However, an infinite time-horizon is not realistic for real-world purposes. In this talk we will look at a couple of metrics which generalise the growth rate for random process that either have more than one long-term behaviour or stop in finite-time.
Week 3 - Intro to decision making with Bayesian Optimization - Juan Ungredda & Paul Kent