BEGIN:VCALENDAR VERSION:2.0 PRODID:talks.ox.ac.uk BEGIN:VEVENT SUMMARY:POSTPONED: Distribution-Free Nonparametric Inference Based on Opti mal Transport and Kernel Methods - Professor Bhaswar B. Bhattacharya (The Wharton School\, University of Pennsylvania) DTSTART;VALUE=DATE-TIME:20260305T153000Z DTEND;VALUE=DATE-TIME:20260305T163000Z UID:https://talks.ox.ac.uk/talks/id/10f16552-474b-4f42-92cc-510c53bfcc63/ DESCRIPTION:The Wilcoxon rank-sum (or Mann–Whitney) test is one of the m ost widely used tools for comparing two groups without making assumptions about the underlying data distribution. One of the reasons for its endurin g popularity is a remarkable result of Hodges and Lehmann (1956)\, which s hows that the asymptotic relative efficiency of Wilcoxon's test with respe ct to Student's t-test\, under location alternatives\, never falls below 0 .864\, despite the former being distribution-free in finite samples. Even more striking is the result of Chernoff and Savage (1958)\, which shows th at the efficiency of a Gaussian score transformed Wilcoxon's test\, agains t the t-test\, is lower bounded by 1. In other words\, the Gaussian score transformed Wilcoxon test uniformly dominates the t-test in terms of effic iency\, while also remaining distribution-free.\n\nIn this talk we will di scuss multivariate versions of these celebrated results\, by considering d istribution-free analogues of the Hotelling T²-test based on optimal tran sport. The proposed tests are consistent against a general class of altern atives and satisfy Hodges-Lehmann and Chernoff-Savage-type efficiency lowe r bounds over various natural families of multivariate distributions\, des pite being entirely agnostic to the underlying data generating mechanism. We will also discuss how optimal-transport-based multivariate ranks can be used to construct distribution-free analogues of the celebrated kernel tw o-sample test that enjoy a trifecta of desirable properties: universal con sistency\, efficient computation\, and nontrivial asymptotic efficiency. \ nSpeakers:\nProfessor Bhaswar B. Bhattacharya (The Wharton School\, Univer sity of Pennsylvania) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/10f16552-474b-4f42-92cc-510c53bfcc63/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:POSTPONED: Distribution-Free Nonparametric Inference Base d on Optimal Transport and Kernel Methods - Professor Bhaswar B. Bhattach arya (The Wharton School\, University of Pennsylvania) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:A Collectivist\, Economic Perspective on AI - Professor Michael I . Jordan (Inria Paris and University of California\, Berkeley) DTSTART;VALUE=DATE-TIME:20260129T153000Z DTEND;VALUE=DATE-TIME:20260129T163000Z UID:https://talks.ox.ac.uk/talks/id/9624b8ec-608c-4444-bbd0-9a02abd1842d/ DESCRIPTION:Information technology is in the midst of a revolution in whic h omnipresent data collection and machine learning are impacting the human world as never before. The word "intelligence" is being used as a North Star for the development of this technology\, with human cognition viewed as a baseline. This view neglects the fact that humans are social animals \, and that much of our intelligence is social and cultural in origin. Th us\, a broader framing is to consider the system level\, where the agents in the system\, be they computers or humans\, are active\, they are cooper ative\, and they wish to obtain value from their participation in learning -based systems. Agents may supply data and other resources to the system o nly if it is in their interest to do so\, and they may be honest and coope rative only if it is in their interest to do so. Critically\, intelligence inheres as much in the overall system as it does in individual agents. T his is a perspective that is familiar in economics\, although without the focus on learning algorithms. A key challenge is thus to bring (micro)eco nomic concepts into contact with foundational issues in the computing and statistical sciences. I'll discuss some concrete examples of problems and solutions at this tripartite interface.\nSpeakers:\nProfessor Michael I. Jordan (Inria Paris and University of California\, Berkeley) LOCATION:Large Lecture Theatre\, Department of Statistics TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/9624b8ec-608c-4444-bbd0-9a02abd1842d/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:A Collectivist\, Economic Perspective on AI - Professor Michael I. Jordan (Inria Paris and University of California\, Berkeley) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Environmental consequences of an ascendant aerospace sector - Prof essor Sebastian Eastham (Imperial College London) DTSTART;VALUE=DATE-TIME:20251029T153000Z DTEND;VALUE=DATE-TIME:20251029T163000Z UID:https://talks.ox.ac.uk/talks/id/7da1d987-9dcb-41da-a3ab-e91a2436e866/ DESCRIPTION:The aerospace sector provides unique and irreplaceable service s but produces outsized environmental impacts which extend far beyond the carbon released by burning jet fuel. As the industry continues to grow exp onentially\, a robust and nuanced understanding of its full climate and ai r quality footprint is essential. This lecture will explore the challengin g methodological problems inherent in quantifying these impacts\, what we can do to improve our understanding\, and possible near-term strategies to achieve clean aviation. \n\nThrough a combination of atmospheric simula tion and analysis of observational data\, we will delve into four specific areas: using deep learning to enable automated detection of condensation trails (contrails)\; the challenge of predicting and addressing contrail c limate impacts\; the global air quality burden of cruise-altitude emission s\; and the nascent field of launch vehicle emissions. These are not just scientific challenges\, but opportunities for rigorous analysis to advance our understanding of the sector's overall environmental footprint.\n\nThi s talk will also discuss how these insights can be leveraged to inform ind ividual choices\, empowering people to reduce their personal environmental footprints\, while also highlighting the coordinated actions required to drive meaningful change across the entire sector. The goal is to show that while this is a profoundly challenging problem\, it is one where scientif ic rigor and innovation are making real progress.\n\nBio: Sebastian Eastha m is the Associate Professor of Sustainable Aviation at Imperial College L ondon\, and a member of the Brahmal Vasudevan Institute for Sustainable Av iation. His research focuses on developing and applying quantitative model s to better understand the environmental effects of the aerospace sector a nd to evaluate strategies for a more sustainable future. With a background in atmospheric science and engineering\, his work provides insights into the complexities of the sector's total environmental footprint\, going bey ond simple carbon emissions to include factors such as contrails and nitro gen oxides. Prior to joining Imperial in 2024 he was Principal Research Sc ientist in MIT's Center for Global Change Science\, and Associate Director of the MIT Laboratory for Aviation and the Environment.\nSpeakers:\nProfe ssor Sebastian Eastham (Imperial College London) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/7da1d987-9dcb-41da-a3ab-e91a2436e866/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Environmental consequences of an ascendant aerospace sect or - Professor Sebastian Eastham (Imperial College London) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Gradient-free stochastic optimization - Professor Alexandre Tsyba kov (CREST-ENSAE Paris) DTSTART;VALUE=DATE-TIME:20250606T150000 DTEND;VALUE=DATE-TIME:20250606T160000 UID:https://talks.ox.ac.uk/talks/id/d99d6716-143d-417a-b843-e703df35b129/ DESCRIPTION:This talk will deal with optimization problems in a statistica l learning setup where the learner has no access to unbiased estimators of the gradient of the objective function. It includes stochastic optimizati on with zero-order oracle\, continuum bandit and contextual continuum ban dit problems. I’ll give an overview of recent results on minimax optimal algorithms and fundamental limits for these problems. \nSpeakers:\nProfes sor Alexandre Tsybakov (CREST-ENSAE Paris) LOCATION:Mathematical Institute (L1\, Mathematical Institute\, Andrew Wile s Building)\, Woodstock Road OX2 6GG TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/d99d6716-143d-417a-b843-e703df35b129/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Gradient-free stochastic optimization - Professor Alexan dre Tsybakov (CREST-ENSAE Paris) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Bayesian Probabilistic Subnational Population Projections - Profes sor Adrian Raftery (University of Washington\, Seattle) DTSTART;VALUE=DATE-TIME:20250616T153000 DTEND;VALUE=DATE-TIME:20250616T163000 UID:https://talks.ox.ac.uk/talks/id/dbab5507-5784-471d-9801-e8ff8b513e86/ DESCRIPTION:Population projections have until recently usually been done d eterministically using the cohort-component method\, yielding a single val ue for each projected future population quantity of interest. Starting in 2015\, the United Nations Population Division changed their approach\, ins tead adopted a fully statistical Bayesian probabilistic approach to projec t fertility\, mortality and population for all countries\, using methods d eveloped by our group. In 2024\, for the first time\, uncertainty about ne t international migration was also included. \n\nIn this approach\, the to tal fertility rate\, female and male life expectancies at birth\, and the net migration rate are projected using Bayesian hierarchical models estima ted via Markov chain Monte Carlo. These are then combined with a cohort-co mponent model\, yielding probabilistic projections for any future quantity of interest. The methodology is implemented in the bayesPop R package\, w hich has been used by the UN to produce the World Population Prospects sin ce 2015. We have recently extended the method to subnational population pr ojections. I will describe the method and some recent extensions\, and ill ustrate it with subnational demographic data from several countries. \nSpe akers:\nProfessor Adrian Raftery (University of Washington\, Seattle) LOCATION:Department of Earth Sciences (Seminar Room\, Department of Earth Sciences)\, South Parks Road OX1 3AN TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/dbab5507-5784-471d-9801-e8ff8b513e86/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Bayesian Probabilistic Subnational Population Projections - Professor Adrian Raftery (University of Washington\, Seattle) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Differentially private M-estimation via noisy optimization - Profe ssor Po-Ling Loh (University of Cambridge) DTSTART;VALUE=DATE-TIME:20250529T151500 DTEND;VALUE=DATE-TIME:20250529T163000 UID:https://talks.ox.ac.uk/talks/id/92f239ab-fced-44e0-ba81-f915ad50733c/ DESCRIPTION:We present a noisy composite gradient descent algorithm for di fferentially private statistical estimation in high dimensions. We begin b y providing general rates of convergence for the parameter error of succes sive iterates under assumptions of local restricted strong convexity and l ocal restricted smoothness. Our analysis is local\, in that it ensures a l inear rate of convergence when the initial iterate lies within a constant- radius region of the true parameter. At each iterate\, multivariate Gaussi an noise is added to the gradient in order to guarantee that the output sa tisfies Gaussian differential privacy. We then derive consequences of our theory for linear regression and mean estimation. Motivated by M-estimator s used in robust statistics\, we study loss functions which downweight the contribution of individual data points in such a way that the sensitivity of function gradients is guaranteed to be bounded\, even without the usua l assumption that our data lie in a bounded domain. We prove that the obje ctive functions thus obtained indeed satisfy the restricted convexity and restricted smoothness conditions required for our general theory. We will also discuss the benefits of acceleration in optimization procedures\, spe cifically a private version of the Frank-Wolfe algorithm\, and its consequ ences for statistical estimation.\n\nThis is based on joint work with Marc o Avella-Medina\, Casey Bradshaw\, Zheng Liu\, and Laurentiu Marchis.\nSpe akers:\nProfessor Po-Ling Loh (University of Cambridge) LOCATION:Mathematical Institute (L3\, Andrew Wiles Building)\, Woodstock R oad OX2 6GG TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/92f239ab-fced-44e0-ba81-f915ad50733c/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Differentially private M-estimation via noisy optimizatio n - Professor Po-Ling Loh (University of Cambridge) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Does AI help humans make better decisions? A statistical evaluatio n framework for experimental and observational studies. - Professor Kosuke Imai (Harvard University) DTSTART;VALUE=DATE-TIME:20250313T160000Z DTEND;VALUE=DATE-TIME:20250313T170000Z UID:https://talks.ox.ac.uk/talks/id/8a6ea51c-ec41-464e-a7b6-d9964ec5bde5/ DESCRIPTION:The use of Artificial Intelligence (AI)\, or more generally da ta-driven algorithms\, has become ubiquitous in today's society. Yet\, in many cases\, and especially when stakes are high\, humans still make final decisions. The critical question\, therefore\, is whether AI helps humans make better decisions compared to a human-alone or AI-alone system. We in troduce a new methodological framework to empirically answer this question with a minimal set of assumptions. We measure a decision maker's ability to make correct decisions using standard classification metrics based on t he baseline potential outcome. We consider a single-blinded and unconfound ed treatment assignment\, where the provision of AI-generated recommendati ons is assumed to be randomized across cases with humans making final deci sions. Under this study design\, we show how to compare the performance of three alternative decision-making systems --- human-alone\, human-with-AI \, and AI-alone. Importantly\, the AI-alone system includes any individual ized treatment assignment\, including those not used in the original study . We also show when AI recommendations should be provided to a human-decis ion maker\, and when one should follow such recommendations. We apply the proposed methodology to our own randomized controlled trial evaluating a p retrial risk assessment instrument. We find that the risk assessment recom mendations do not improve the classification accuracy of a judge's decisio n to impose cash bail. Furthermore\, we find that replacing a human judge with algorithms --- the risk assessment score and a large language model i n particular --- leads to a worse classification performance.\nSpeakers:\n Professor Kosuke Imai (Harvard University) LOCATION:Seminar Room\, Department of Earth Sciences TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/8a6ea51c-ec41-464e-a7b6-d9964ec5bde5/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Does AI help humans make better decisions? A statistical evaluation framework for experimental and observational studies. - Profess or Kosuke Imai (Harvard University) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:E-Values\, Anytime-Validity and Bayes - Professor Peter Grunwald ( Leiden University) DTSTART;VALUE=DATE-TIME:20250220T160000Z DTEND;VALUE=DATE-TIME:20250220T170000Z UID:https://talks.ox.ac.uk/talks/id/b2ad754a-3bf7-415d-b39e-864d382ddf55/ DESCRIPTION:E-values (wikipedia) are an alternative to p-values that effor tlessly deal with optional continuation: with e-value based tests and the corresponding anytime valid (AV) confidence intervals\, one can always gat her additional data\, while keeping statistically valid conclusions. Until 2019\, publications on e-values were few and far between: the concept did not even have a name. Then\, in the course of a few months\, four papers by different research groups\, (including ours - see below) appeared on ar Xiv that firmly established them as an important statistical concept. By n ow\, there are 100s of papers on e-values and there have been two internat ional workshops on the topic. Allowing for optional continuation is just o ne way in which e-values provide more flexibility than p-values – they also allow to set a type of significance/confidence level alpha after see ing the data\, which is a mortal sin in classical testing. In this talk I will introduce e-values\, e-processes and AV confidence intervals\, and di scuss how like Bayesian approaches\, they employ priors\, while\, unlike i n Bayesian approaches\, we obtain error guarantees even if these priors mi salign with the data. \n \nMain literature: \nG.\, De Heide\, Koolen. Safe Testing. Journal of the Royal Statistical Society Series B\, 2024 (first version appeared on arXiv 2019).\nG. Beyond Neyman-Pearson: e-values enab le hypothesis testing with a data-driven alpha. Proceedings National Acade my of Sciences of the USA (PNAS)\, 2024.\n\nSpeakers:\nProfessor Peter Gru nwald (Leiden University) LOCATION:Seminar Room\, Department of Earth Sciences TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/b2ad754a-3bf7-415d-b39e-864d382ddf55/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:E-Values\, Anytime-Validity and Bayes - Professor Peter G runwald (Leiden University) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:The trajectories of complex disease - Professor Xavier Didelot (Un iversity of Warwick) DTSTART;VALUE=DATE-TIME:20250206T160000Z DTEND;VALUE=DATE-TIME:20250206T170000Z UID:https://talks.ox.ac.uk/talks/id/bc350677-d539-4c2d-b7ac-48f630eeb2fd/ DESCRIPTION:Genomic data are increasingly being used to understand infecti ous disease epidemiology. Isolates from a given outbreak are sequenced\, a nd the patterns of shared variation are used to infer phylogenetic trees. However these are not directly informative about who infected whom: a phyl ogenetic tree is not a transmission tree. A transmission tree can be infer red from a phylogeny while accounting for within-host genetic diversity by coloring the branches of a phylogeny according to which host those branch es were in. We show how this approach can be applied to partially sampled and ongoing outbreaks. This requires computing the correct probability of a partially observed transmission tree and we demonstrate how to do this f or a large class of epidemiological models. The resulting uncertainty on w ho infected whom can be high and we explore two solutions to this problem: the use of several genomes per host\, and the use of additional epidemiol ogical data.\nSpeakers:\nProfessor Xavier Didelot (University of Warwick) LOCATION:Seminar Room\, Department of Earth Sciences TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/bc350677-d539-4c2d-b7ac-48f630eeb2fd/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:The trajectories of complex disease - Professor Xavier Di delot (University of Warwick) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:The trajectories of complex disease - Professor Gil McVean (The El lison Institute of Technology\, Oxford) DTSTART;VALUE=DATE-TIME:20241024T153000 DTEND;VALUE=DATE-TIME:20241024T163000 UID:https://talks.ox.ac.uk/talks/id/2079adab-7381-434a-8432-5d8d056c6012/ DESCRIPTION:The analysis of longitudinal data from electronic health recor ds (EHRs) has the potential to improve clinical diagnosis and enable perso nalised medicine\, motivating efforts to identify disease commonalities an d subtypes from patient comorbidity information and other modalities. We h ave developed an age-dependent topic-modelling (ATM) method that provides a low-rank representation of longitudinal records of hundreds of distinct diseases in large EHR datasets and applied it to c. 300\,000 individuals f rom UK Biobank and >200\,000 individuals from the All of Us program. A sur prisingly small number of disease trajectories capture known and novel com binations of disorders that occur throughout life and identify disease sub types that occur in multiple topics\, with differential genetic risk profi les. Such stratification improves understanding of patient risk and hetero geneity\, leading to better identification of genetic risk\, characterisat ion of pathological pathways and the discovery of new therapeutic targets. \nSpeakers:\nProfessor Gil McVean (The Ellison Institute of Technology\, O xford) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/2079adab-7381-434a-8432-5d8d056c6012/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:The trajectories of complex disease - Professor Gil McVea n (The Ellison Institute of Technology\, Oxford) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Approximate gradients for inference of partially-observed stochast ic processes - Professor Marc Suchard (David Geffen School of Medicine and Department of Biostatistics at UCLA) DTSTART;VALUE=DATE-TIME:20240507T153000 DTEND;VALUE=DATE-TIME:20240507T163000 UID:https://talks.ox.ac.uk/talks/id/cfef9751-30c0-4c9f-ad30-56ae6298180c/ DESCRIPTION:Bayesian computation remains onerous at scale for inference un der many discrete-valued stochastic process-based models\, while these mod els remain ubiquitous across biology and public health. In this talk\, we will explore how one can construct computationally efficient approximatio ns to the gradient of the data likelihood under continuous-time Markov cha in (CTMC) models with respect to their high-dimensional parameters. CTMCs underpin the most popular models for learning about how rapidly evolving pathogens change over time and space to give rise to human infection\, and the dimensionality of these problems are daunting. With these approximati ons in hand\, a new variant of Hamiltonian Monte Carlo (HMC) becomes tract able to explore the parameter posterior\, and we bound the approximation e rror using several small tricks from matrix analysis. This new sampling a pproach enables the introduction of a novel random-effects CTMC model that captures biological realism previously missing. Applied to the analysis of early SARS-CoV-2 genomes\, the random-effects remove bias in inference of the location and timing of the pathogen's split-over into humans\, whil e the approximate-gradient-based machinery is over an order of magnitude m ore time efficient than conventional sampling approaches.\nSpeakers:\nProf essor Marc Suchard (David Geffen School of Medicine and Department of Bios tatistics at UCLA) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/cfef9751-30c0-4c9f-ad30-56ae6298180c/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Approximate gradients for inference of partially-observed stochastic processes - Professor Marc Suchard (David Geffen School of Med icine and Department of Biostatistics at UCLA) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Tackling the hidden costs of computational science: GREENER princi ples for environmentally sustainable research - Dr Loïc Lannelongue (Hear t and Lung Research Institute\, University of Cambridge and the Cambridge- Baker Systems Genomics Initiative) DTSTART;VALUE=DATE-TIME:20240426T153000 DTEND;VALUE=DATE-TIME:20240426T163000 UID:https://talks.ox.ac.uk/talks/id/07d4447f-9324-4380-b700-dafdac60cfbc/ DESCRIPTION:From genetic studies and astrophysics simulations to statistic al modelling and AI\, scientific computing has enabled amazing discoveries and there is no doubt it will continue to do so. However\, the correspond ing environmental impact is a growing concern in light of the urgency of t he climate crisis\, so what can we all do about it? Tackling this issue an d making it easier for scientists to engage with sustainable computing is what motivated the Green Algorithms project. Through the prism of the GREE NER principles for environmentally sustainable science\, we will discuss w hat we learned along the way\, how to estimate the impact of our work and what levers scientists and institutions have to make their research more s ustainable. We will also debate what hurdles exist and what is still neede d moving forward.\nSpeakers:\nDr Loïc Lannelongue (Heart and Lung Researc h Institute\, University of Cambridge and the Cambridge-Baker Systems Geno mics Initiative) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/07d4447f-9324-4380-b700-dafdac60cfbc/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Tackling the hidden costs of computational science: GREEN ER principles for environmentally sustainable research - Dr Loïc Lannelon gue (Heart and Lung Research Institute\, University of Cambridge and the C ambridge-Baker Systems Genomics Initiative) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Statistics\, nursing\, and social reform: Following in the footste ps of Florence Nightingale - Professor Barbara Engelhardt (Senior Investig ator at Gladstone Institutes and Professor at Stanford University in the D epartment of Biomedical Data Science) DTSTART;VALUE=DATE-TIME:20240223T153000Z DTEND;VALUE=DATE-TIME:20240223T163000Z UID:https://talks.ox.ac.uk/talks/id/fa010ec7-c6f7-46b3-b5af-26f6f05a5968/ DESCRIPTION:Decision-making tasks in healthcare settings use methods that make a number of assumptions that we know are violated in clinical data. F or example\, clinicians do not always act optimally\; clinicians are more or less aggressive in treating patients\; clinicians have biases\; and pat ients have (often unobserved) conditions that lead to differential respons e to interventions. In this talk\, and following in Florence Nightingale's path\, I will walk through a handful of these violated assumptions and di scuss statistical reinforcement learning and inverse reinforcement learnin g methods to address these violated assumptions. I will show on a number o f scenarios\, including sepsis treatment and electrolyte repletion\, that these methods that have more flexible assumptions than existing methods le ad to substantial improvements in decision-making tasks in clinical settin gs\, reducing bias and leading to improved clinical outcomes.\nSpeakers:\n Professor Barbara Engelhardt (Senior Investigator at Gladstone Institutes and Professor at Stanford University in the Department of Biomedical Data Science) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/fa010ec7-c6f7-46b3-b5af-26f6f05a5968/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Statistics\, nursing\, and social reform: Following in th e footsteps of Florence Nightingale - Professor Barbara Engelhardt (Senior Investigator at Gladstone Institutes and Professor at Stanford University in the Department of Biomedical Data Science) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Introducing the Forster-Warmuth Nonparametric Counterfactual Regre ssion - Professor Eric Tchetgen Tchetgen (University of Pennsylvania) DTSTART;VALUE=DATE-TIME:20231026T153000 DTEND;VALUE=DATE-TIME:20231026T163000 UID:https://talks.ox.ac.uk/talks/id/a6d1ed9f-bdf1-4965-b901-1f59e49bb34e/ DESCRIPTION:Series regression estimation is one of the most popular non-pa rametric regression techniques in practice. The most routinely used series estimator is based on ordinary least squares fitting\, which is known to be minimax rate optimal in various settings\, albeit under stringent restr ictions on the basis functions. In this work\, inspired by the recently de veloped Forster-Warmuth (FW) regression\, we propose an alternative nonpar ametric series estimator that can attain minimax estimation rates under st rictly weaker conditions imposed on the basis functions\, than virtually a ll existing series estimators in the literature. Another contribution of t his work generalizes the FW-regression to a so-called counterfactual regre ssion problem\, in which the response variable of interest may not be dire ctly observed (hence\, the name ``counterfactual'') on all sampled units. Although counterfactual regression is not entirely a new area of inquiry\, we propose the first-ever systematic study of this challenging problem fr om a unified pseudo-outcome perspective. In fact\, we provide what appears to be the first generic and constructive approach for generating the pseu do-outcome (to substitute for the unobserved response) which leads to the estimation of the counterfactual regression curve of interest with small b ias\, namely bias of second order. Several applications are used to illust rate the resulting FW counterfactual regression including a large class of nonparametric regression problems in missing data and causal inference li terature\, for which we establish conditions for minimax rate optimality. This is joint work with Yachong Yang and Arun kuchibhotla.\nSpeakers:\nPro fessor Eric Tchetgen Tchetgen (University of Pennsylvania) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/a6d1ed9f-bdf1-4965-b901-1f59e49bb34e/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Introducing the Forster-Warmuth Nonparametric Counterfact ual Regression - Professor Eric Tchetgen Tchetgen (University of Pennsylva nia) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Understanding neural networks and quantification of their uncertai nty via exactly solvable models - Professor Lenka Lenka Zdeborová (École Polytechnique Fédérale de Lausanne) DTSTART;VALUE=DATE-TIME:20230505T153000 DTEND;VALUE=DATE-TIME:20230505T163000 UID:https://talks.ox.ac.uk/talks/id/cde53ef6-143e-4988-889f-7f05d1da1d88/ DESCRIPTION:The affinity between statistical physics and machine learning has a long history. Theoretical physics often proceeds in terms of solvabl e synthetic models\; I will describe the related line of work on solvable models of simple feed-forward neural networks. I will then discuss how thi s approach allows us to analyze uncertainty quantification in neural netwo rks\, a topic that gained urgency in the dawn of widely deployed artificia l intelligence. I will conclude with what I perceive as important specific open questions in the field. \nSpeakers:\nProfessor Lenka Lenka Zdeborov á (École Polytechnique Fédérale de Lausanne) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/cde53ef6-143e-4988-889f-7f05d1da1d88/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Understanding neural networks and quantification of their uncertainty via exactly solvable models - Professor Lenka Lenka Zdeborov á (École Polytechnique Fédérale de Lausanne) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Causal learning from observational data - Professor Marloes Maathu is (ETH\, Zurich) DTSTART;VALUE=DATE-TIME:20230203T153000Z DTEND;VALUE=DATE-TIME:20230203T163000Z UID:https://talks.ox.ac.uk/talks/id/50442fbc-b3d6-4291-a2a3-28a9c16339bd/ DESCRIPTION:I will discuss a line of work on estimating causal effects fro m observational data. In the first part of the talk\, I will discuss ident ification and estimation of causal effects when the underlying causal grap h is known\, using adjustment. In the second part\, I will discuss what on e can do when the causal graph is unknown. Throughout\, examples will be u sed to illustrate the concepts and no background in causality is assumed. \nSpeakers:\nProfessor Marloes Maathuis (ETH\, Zurich) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/50442fbc-b3d6-4291-a2a3-28a9c16339bd/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Causal learning from observational data - Professor Marlo es Maathuis (ETH\, Zurich) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Compositional data analysis: A fresh approach - Professor David F irth (University of Warwick) DTSTART;VALUE=DATE-TIME:20230222T153000Z DTEND;VALUE=DATE-TIME:20230222T163000Z UID:https://talks.ox.ac.uk/talks/id/90bb5f33-8f8b-4095-baa5-2a413e5fbae5/ DESCRIPTION:The analysis of composition is a highly active area of applied multivariate statistics\, in fields that span physical and biological sci ence\, the social sciences and humanities. The composition of a multi-par t entity is a fixed-sum vector representing the relative amounts of each p art in the whole entity. Just a few examples are: in electoral politics\ , the vote shares for different political parties\; in geology\, the perce ntages of different minerals in rock samples\; in sociology and other beha vioural sciences\, the "time budgets" of individuals\, e.g.\, fractions of the day that are spent sleeping\, working\, sedentary\, physically active \, etc.\; in biology\, the relative prevalence of different microbes in th e human gut.\n\nThe statistical literature in this area is dominated by wo rk done in the 1980s by John Aitchison\, including the highly influential 1986 book "The Statistical Analysis of Compositional Data". The essence o f Aitchison's approach is to transform composition vectors to a set of con trasts among logarithms of the data\, and then work with the standard tool s of multivariate statistics such as multivariate normal distributions and linear models. In this talk I challenge that approach\, in terms of both its foundational principles and some well-known practical difficulties. A more flexible approach\, based on suitably targeted statistical models r ather than on data-transformation\, is advocated.\nSpeakers:\nProfessor Da vid Firth (University of Warwick) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/90bb5f33-8f8b-4095-baa5-2a413e5fbae5/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Compositional data analysis: A fresh approach - Professo r David Firth (University of Warwick) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:From population to person: Counterfactual risk prediction - Profe ssor Ruth Keogh (London School of Hygiene & Tropical Medicine) DTSTART;VALUE=DATE-TIME:20221201T153000Z DTEND;VALUE=DATE-TIME:20221201T163000Z UID:https://talks.ox.ac.uk/talks/id/c2de1639-47e2-4530-a3cd-d7b7ac04aaad/ DESCRIPTION:Clinical risk prediction models enable predictions of a person ’s risk of an outcome (e.g. mortality) given their observed characterist ics. It is often of interest to use risk predictions to inform whether a p erson should initiate a particular treatment. However\, when standard clin ical prediction models are developed in a population in which patients fol low a mix of treatment strategies\, they are unsuitable for informing trea tment decisions. Counterfactual risk predictions aim to address this probl em - they are estimates of what a person’s risk would be if they were to follow a particular treatment strategy\, given their individual character istics that are also predictive of the outcome. \nCausal inference methods typically focus on estimating population average treatment effects. In th is talk I will discuss how causal inference methods can be used for indivi dual counterfactual risk prediction using longitudinal observational data on treatment use\, patient characteristics and a time-to-event outcome. An essential step in development and reporting of prediction models is to va lidate their performance. I will discuss the challenges of this\, and desc ribe some new methods for assessing the predictive performance of counterf actual risk prediction.\n\nIn a motivating example\, we are interested in counterfactual risk predictions for mortality in patients awaiting a liver transplant under the strategies of receiving or not receiving a transplan t. I will illustrate the methods using data from the US Scientific Registr y of Transplant Patients.\n\nSpeakers:\nProfessor Ruth Keogh (London Schoo l of Hygiene & Tropical Medicine) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/c2de1639-47e2-4530-a3cd-d7b7ac04aaad/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:From population to person: Counterfactual risk prediction - Professor Ruth Keogh (London School of Hygiene & Tropical Medicine) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:A path to personalised disease prevention: using genomics to predi ct risk for common diseases - Peter Donnelly (CEO of Genomics PLC and Prof essor of Statistical Science at University of Oxford) DTSTART;VALUE=DATE-TIME:20221021T153000 DTEND;VALUE=DATE-TIME:20221021T163000 UID:https://talks.ox.ac.uk/talks/id/f637bdfd-d154-49c7-bcf9-607a85794557/ DESCRIPTION:It has long been known that genetics is a major risk factor fo r all the common chronic human diseases\, such as heart disease\, diabetes \, osteoporosis\, and the mental health and auto-immune disorders\, and fo r the common cancers\, such as breast\, prostate\, and bowel cancer. For m any of these diseases\, or in some cases for important subsets of individu als\, it is the single most important predictor of disease risk. We now kn ow\, from 20 years of human genetics studies\, that this risk is polygenic : for any given disease\, it is due to the cumulative effects of a very la rge number of different genetic variants in our DNA\, each of individually small impact. Polygenic risk scores (PRSs) provide a single measure whic h quantifies the overall impact of these variants. For many diseases the u nderlying genetic studies are now large enough to provide good information about which variants matter for a particular disease\, and their effect s izes\, and in parallel we now have large prospective population cohorts\, such as UK Biobank\, in which to assess and validate predictive power. \n\ nThe talk will describe our work in Genomics plc on PRSs and their use\, t ypically in combination with non-genetic risk factors\, to provide a new l evel of risk prediction for common diseases\, and the path to their adopti on in healthcare systems to empower a new generation of risk-stratified\, personalised\, disease prevention. The identification of individuals at in creased risk of disease who are currently invisible to health systems will allow those systems to get more of the right individuals into the appropr iate screening\, prevention\, and treatment pathways. For the individual\, this can help to prevent disease entirely or to catch it early\, when out comes are much better\; for the health system\, it means existing preventi on and screening resources are used more efficiently by deploying them on higher-risk individuals\; and from a population health point of view\, by moving healthcare towards prevention\, lives and resources are saved over time.\n\nSpeakers:\nPeter Donnelly (CEO of Genomics PLC and Professor of S tatistical Science at University of Oxford) LOCATION:24-29 St Giles' (Large Lecture Theatre)\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/f637bdfd-d154-49c7-bcf9-607a85794557/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:A path to personalised disease prevention: using genomics to predict risk for common diseases - Peter Donnelly (CEO of Genomics PLC and Professor of Statistical Science at University of Oxford) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Alan Turing at 110: current questions and issues - Professor Andre w Hodges (University of Oxford) DTSTART;VALUE=DATE-TIME:20220623T140000 DTEND;VALUE=DATE-TIME:20220623T150000 UID:https://talks.ox.ac.uk/talks/id/6f9c688d-a726-4267-8089-4597e7ce7274/ DESCRIPTION:This talk marks the 110th anniversary of Alan Turing’s birth . Only a very few people alive today can remember him\, but the questions raised by his life and work are as lively as ever. I will range over a var iety of topics\, centred on questions about computability and artificial i ntelligence\, but also reflecting the very wide range of Turing’s contri butions to science and history.\n\n\nSpeakers:\nProfessor Andrew Hodges (U niversity of Oxford) LOCATION:24-29 St Giles\, Oxford TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/6f9c688d-a726-4267-8089-4597e7ce7274/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Alan Turing at 110: current questions and issues - Profes sor Andrew Hodges (University of Oxford) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Ethics from the perspective of an applied statistician - Professo r Denise Lievesley (Honorary Fellow of Green Templeton College\, Universit y of Oxford) DTSTART;VALUE=DATE-TIME:20220210T153000Z DTEND;VALUE=DATE-TIME:20220210T163000Z UID:https://talks.ox.ac.uk/talks/id/f8f900a2-65e4-458f-ab71-760d0a6f5d84/ DESCRIPTION:Statisticians work in a wide variety of different political an d cultural environments which influence their autonomy and their status\, which in turn impact on the ethical frameworks they employ. The need for a UN-led fundamental set of principles governing official statistics becam e apparent at the end of the 1980s when countries in Central Europe began to change from centrally planned economies to market-oriented democracies. It was essential to ensure that national statistical systems in such coun tries would be able to produce appropriate and reliable data that adhered to certain professional and scientific standards. Alongside the UN initia tive\, a number of professional statistical societies adopted codes of con duct. \n\nDo such sets of principles and ethical codes remain relevant ove r time? Or do changes in the way statistics are compiled and used mean th at we need to review and adapt them? For example as combining data source s becomes more prevalent\, record linkage\, in particular\, poses privac y and ethical challenges. Similarly obtaining informed consent from units for access to and linkage of their data from non-survey sources continues to be challenging. Denise will particularly draw on her earlier role as a statistician in the United Nations\, working with some 200 countries\, to discuss some of the ethical issues she encountered then and how these mig ht change over time. \nSpeakers:\nProfessor Denise Lievesley (Honorary Fe llow of Green Templeton College\, University of Oxford) LOCATION:24-29 St Giles' (Large Lecture Theatre)\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/f8f900a2-65e4-458f-ab71-760d0a6f5d84/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Ethics from the perspective of an applied statistician - Professor Denise Lievesley (Honorary Fellow of Green Templeton College\, University of Oxford) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Modelling infectious diseases: what can branching processes tell u s? - Samir Bhatt\, Professor of Machine Learning and Public Health (Univer sity of Copenhagen and Professor of Statistics and Public Health\, Imperia l College London) DTSTART;VALUE=DATE-TIME:20220224T153000Z DTEND;VALUE=DATE-TIME:20220224T163000Z UID:https://talks.ox.ac.uk/talks/id/aeda56d0-a4cd-48f7-a675-72ce2c661c91/ DESCRIPTION:Mathematical descriptions of infectious disease outbreaks are fundamental to understanding how transmission occurs. Reductively\, two ap proaches are used: individual based simulators and governing equation mode ls\, and both approaches have a multitude of pros and cons. In this talk I will connect these two worlds via general branching processes. I will dis cuss (at a high level) the rather beautiful mathematics that arises from t hese branching processes and how these can help us understand the assumpti ons underpinning mathematical models for infectious disease. I will then e xplain how this new maths can help us understand uncertainty better\, and show some simple examples. This talk will be a little technical\, but I wi ll focus as much as possible on intuition and the big picture.\nSpeakers:\ nSamir Bhatt\, Professor of Machine Learning and Public Health (University of Copenhagen and Professor of Statistics and Public Health\, Imperial Co llege London) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/aeda56d0-a4cd-48f7-a675-72ce2c661c91/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Modelling infectious diseases: what can branching process es tell us? - Samir Bhatt\, Professor of Machine Learning and Public Healt h (University of Copenhagen and Professor of Statistics and Public Health\ , Imperial College London) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY: Statistics and the fight against modern slavery - Professor Sir B ernard Silverman (University of Nottingham and University of Oxford) DTSTART;VALUE=DATE-TIME:20220304T150000Z DTEND;VALUE=DATE-TIME:20220304T170000Z UID:https://talks.ox.ac.uk/talks/id/70fdc9e9-86c3-4da3-9e6d-39c89878e917/ DESCRIPTION:Modern slavery takes many forms and has only in recent decades come to wide public and political consciousness. The Modern Slavery Act 2015 consolidated a number of relevant offences\, and introduced a maximum penalty of life imprisonment. The numbers of victims known to the author ities\, and of prosecutions\, have considerably increased in recent years\ , but modern slavery persists\, both in the UK and worldwide.\n\nA number of different statistical approaches have been used in the fight against m odern slavery and other hidden crimes. I will survey and explore some of these\, and consider their role leading up to the Modern Slavery Act and i n subsequent developments.\n\nOverall I hope that the lecture will provide the opportunity to reflect more widely on what is the best appropriate st atistical contribution to what is an important and disturbing aspect of pu blic policy. In particular there are broader questions about presenting s tatistical evidence in an area where there is so much uncertainty. Should we follow George Box’s dictum that “all models are wrong\, but some a re useful”? There are some (imperfect) analogies with Florence Nighting ale’s own work and campaign leading to the Public Health Act 1875\, whic h saved the lives of millions.\nSpeakers:\nProfessor Sir Bernard Silverman (University of Nottingham and University of Oxford) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/70fdc9e9-86c3-4da3-9e6d-39c89878e917/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk: Statistics and the fight against modern slavery - Profes sor Sir Bernard Silverman (University of Nottingham and University of Oxfo rd) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Highly accurate protein structure prediction with AlphaFold - Dr J ohn Jumper (DeepMind) DTSTART;VALUE=DATE-TIME:20211028T153000 DTEND;VALUE=DATE-TIME:20211028T163000 UID:https://talks.ox.ac.uk/talks/id/f9774caf-2d0a-461e-9323-23070e195bc4/ DESCRIPTION:Predicting a protein’s structure from its primary sequence h as been a grand challenge in biology for the past 50 years\, holding the p romise to bridge the gap between the pace of genomics discovery and result ing structural characterization. In this talk\, we will describe work at D eepMind to develop AlphaFold\, a new deep learning-based system for struct ure prediction that achieves high accuracy across a wide range of targets. We demonstrated our system in the 14th biennial Critical Assessment of P rotein Structure Prediction (CASP14) across a wide range of difficult targ ets\, where the assessors judged our predictions to be at an accuracy “c ompetitive with experiment” for approximately 2/3rds of proteins. The ta lk will cover both the underlying machine learning ideas and the implicati ons for biological research.\nSpeakers:\nDr John Jumper (DeepMind) LOCATION:24-29 St Giles'\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/f9774caf-2d0a-461e-9323-23070e195bc4/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Highly accurate protein structure prediction with AlphaFo ld - Dr John Jumper (DeepMind) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Assessing Personalization in Digital Health - Professor Susan Murp hy (Harvard University) DTSTART;VALUE=DATE-TIME:20210618T153000 DTEND;VALUE=DATE-TIME:20210618T163000 UID:https://talks.ox.ac.uk/talks/id/dc140ede-22b6-484f-88b8-42202c3e775e/ DESCRIPTION:Reinforcement Learning provides an attractive suite of online learning methods for personalizing interventions in a Digital Health. H owever after an reinforcement learning algorithm has been run in a clinica l study\, how do we assess whether personalization occurred? We might fin d users for whom it appears that the algorithm has indeed learned in which contexts the user is more responsive to a particular intervention. But c ould this have happened completely by chance? We discuss some first appr oaches to addressing these questions.\nSpeakers:\nProfessor Susan Murphy ( Harvard University) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/dc140ede-22b6-484f-88b8-42202c3e775e/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Assessing Personalization in Digital Health - Professor S usan Murphy (Harvard University) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:On classification with small Bayes error and the max-margin classi fier - Professor Sara Van de Geer (ETH Zurich) DTSTART;VALUE=DATE-TIME:20210429T153000 DTEND;VALUE=DATE-TIME:20210429T163000 UID:https://talks.ox.ac.uk/talks/id/02bef1f8-c7fc-42a3-880b-27342ae1f28a/ DESCRIPTION:This is joint work with Geoffrey Chinot\, Felix Kuchelmeister and Matthias Löffler.\n\nWe consider the classification problem where one observes a design matrix X ∈ Rn×p and a binary response variable Y = s ign(Xβ* +ξ) ∈ {±1}n. Here β* ∈ Rp is an vector of unknown coeffici ents with llβ*ll2= 1 and ξ ∼ N(0\, σ2I) ∈ Rn is an unobservable noi se vector independent of X. We will present some theoretical results on th e misclassification error of a class of estimators β of β* which are bas ed on L1-regularization or L1-interpolation. The emphasis in this talk wil l be on the interpolating estimator. It is observed in empirical studies t hat classification algorithms achieving zero training error can perform we ll in test sets. We aim at contributing to a theoretical understanding of this phenomenon in the high-dimensional situation (i.e. p >> n). To allow for small test error we focus on the case where σ2 is small. In the speci al setting of i.i.d. Gaussian design\, we examine the minimum L1-norm inte rpolator or max-margin classifier and its rate of convergence under L1-spa rsity assumptions. Related is the noisy one-bit compressed sensing problem \, where we apply the algorithm of Plan and Vershynin [2013] and (re-)esta blish rates under L0– and L1-sparsity conditions.\n\nReferences\nY. Plan and R. Vershynin. One-bit compressed sensing by linear programming\nCommu nications on Pure and Applied Mathematics\, 66(8):1275–1297\, 2013.\nSpe akers:\nProfessor Sara Van de Geer (ETH Zurich) LOCATION:24-29 St Giles'\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/02bef1f8-c7fc-42a3-880b-27342ae1f28a/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:On classification with small Bayes error and the max-marg in classifier - Professor Sara Van de Geer (ETH Zurich) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Finding Today’s Slaves: Lessons Learned From Over A Decade of Me asurement in Modern Slavery - Professor Davina Durgana (American Universit y\, Washington\, DC) DTSTART;VALUE=DATE-TIME:20210225T153000Z DTEND;VALUE=DATE-TIME:20210225T163000Z UID:https://talks.ox.ac.uk/talks/id/3e1e78e0-ca01-43f4-bc97-5f0b6834fb4a/ DESCRIPTION:Dr. Durgana will present her work leading and innovating the u se of statistics in the global modern slavery vulnerability and prevalence field over the past decade. She will present work on the Global Estimates of Modern Slavery with the United Nations\, Global Slavery Index\, and on application of Multiple Systems Estimation throughout Europe with the UN Office on Drugs and Crime. She will also discuss compelling developments w ithin leading national governments on prevalence estimation and how her wo rk engages with the global policy community.\nSpeakers:\nProfessor Davina Durgana (American University\, Washington\, DC) LOCATION:24-29 St Giles'\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/3e1e78e0-ca01-43f4-bc97-5f0b6834fb4a/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Finding Today’s Slaves: Lessons Learned From Over A Dec ade of Measurement in Modern Slavery - Professor Davina Durgana (American University\, Washington\, DC) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Veridical Data Science for biomedical discovery: detecting epistat ic interactions with epiTree - Professor Bin Yu (UC Berkeley) DTSTART;VALUE=DATE-TIME:20210218T153000Z DTEND;VALUE=DATE-TIME:20210218T163000Z UID:https://talks.ox.ac.uk/talks/id/1a65cc08-dd84-441c-80be-13ab54f4b007/ DESCRIPTION:“A.I. is like nuclear energy — both promising and dangerou s” — Bill Gates\, 2019.\n\nData Science is a pillar of A.I. and has dr iven most of recent cutting-edge discoveries in biomedical research. In pr actice\, Data Science has a life cycle (DSLC) that includes problem formul ation\, data collection\, data cleaning\, modeling\, result interpretation and the drawing of conclusions. Human judgement calls are ubiquitous at e very step of this process\, e.g.\, in choosing data cleaning methods\, pre dictive algorithms and data perturbations. Such judgment calls are often r esponsible for the “dangers” of A.I. To maximally mitigate these dange rs\, we developed a framework based on three core principles: Predictabili ty\, Computability and Stability (PCS). Through a workflow and documentati on (in R Markdown or Jupyter Notebook) that allows one to manage the whole DSLC\, the PCS framework unifies\, streamlines and expands on the best pr actices of machine learning and statistics – bringing us a step forward towards veridical Data Science.\nIn this lecture\, we will illustrate the PCS framework through the epiTree\; a pipeline to discover epistasis inter actions from genomics data. epiTree addresses issues of scaling of penetra nce through decision trees\, significance calling through PCS p-values\, a nd combinatorial search over interactions through iterative random forests (which is a special case of PCS). Using UK Biobank data\, we validate th e epiTree pipeline through an application to the red-hair phenotype\, wher e several genes are known to display epistatic interactions.\nSpeakers:\nP rofessor Bin Yu (UC Berkeley) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/1a65cc08-dd84-441c-80be-13ab54f4b007/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Veridical Data Science for biomedical discovery: detectin g epistatic interactions with epiTree - Professor Bin Yu (UC Berkeley) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:(Not) Aggregating Data - Professor Kerrie Mengersen (Queensland Un iversity of Technology in the Science and Engineering Faculty) DTSTART;VALUE=DATE-TIME:20210121T100000Z DTEND;VALUE=DATE-TIME:20210121T110000Z UID:https://talks.ox.ac.uk/talks/id/36e8605e-5a6a-4da6-ba66-ecfce9454d92/ DESCRIPTION:The ability to generate\, access and combine multiple sources of data presents both opportunity and challenge for statistical science. A n exemplar phenomenon is the charge to collate all relevant data for the p urposes of comprehensive control and analysis. However\, this ambition is often thwarted by the relentless expansion in volume of data\, as well as issues of data provenance\, privacy and governance. Alternatives to creati ng ‘the one database to rule them all’ are emerging. An appealing appr oach is the concept of federated learning\, also known as distributed anal ysis\, which aims to analyse disparate datasets in situ. In this presentat ion\, I will discuss some case studies that have motivated our interest in federated learning\, review the statistical and computational issues invo lved in the development of such an approach\, and outline our recent effor ts to understand and implement a federated learning model in the context o f the Australian Cancer Atlas.\nSpeakers:\nProfessor Kerrie Mengersen (Que ensland University of Technology in the Science and Engineering Faculty) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/36e8605e-5a6a-4da6-ba66-ecfce9454d92/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:(Not) Aggregating Data - Professor Kerrie Mengersen (Quee nsland University of Technology in the Science and Engineering Faculty) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Florence Nightingale and the politicians’ pigeon holes: using da ta for the good of society - Professor Deborah Ashby (Royal Statistical So ciety) DTSTART;VALUE=DATE-TIME:20201201T150000Z DTEND;VALUE=DATE-TIME:20201201T160000Z UID:https://talks.ox.ac.uk/talks/id/edf76c8e-bfc6-4057-913c-aee2cbb10a43/ DESCRIPTION:Florence Nightingale\, best known as the Lady with the Lamp\, is recognised as a pioneering and passionate statistician. She was also pa ssionate about education\, having argued successfully with her parents to be allowed to study mathematics\, and later nursing\, herself. More wide ly\, she offered opinions on the education of children\, soldiers\, army d octors\, and nurses\, as well as railing against the ‘enforced idleness ’ of women. A particular concern was the lack of statistical literacy am ong politicians. As we celebrate the bicentenary of her birth\, the need f or education in statistical and data skills shows no signs of abating. Wha t advice would Florence Nightingale offer were she here today?\nSpeakers:\ nProfessor Deborah Ashby (Royal Statistical Society) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/edf76c8e-bfc6-4057-913c-aee2cbb10a43/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Florence Nightingale and the politicians’ pigeon holes: using data for the good of society - Professor Deborah Ashby (Royal Stati stical Society) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Exploring the Data Visualizations of W.E.B. Du Bois - Jason Forres t (McKinsey & Co\, New York) DTSTART;VALUE=DATE-TIME:20201023T153000 DTEND;VALUE=DATE-TIME:20201023T164500 UID:https://talks.ox.ac.uk/talks/id/f5e99a0c-5d87-453b-a1c2-307d5dc307cb/ DESCRIPTION:At the 1900 Paris Exposition\, an all African-American team le ad by scholar and activist W.E.B. Du Bois sought to challenge and recontex tualize the perception of African-Americans at the dawn of the 20th-centur y. In less than 5 months\, his team conducted sociological research and ha nd-made more than 60 large data visualization posters for a massive Europe an audience which ultimately awarded Du Bois a gold medal for his efforts. While relatively obscure until recently\, the ramification of his landmar k work remains challenging and especially important in light of the Black Lives Matter movement.\nSpeakers:\nJason Forrest (McKinsey & Co\, New York ) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/f5e99a0c-5d87-453b-a1c2-307d5dc307cb/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Exploring the Data Visualizations of W.E.B. Du Bois - Jas on Forrest (McKinsey & Co\, New York) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:(POSTPONED) Exploring the Data Visualizations of W.E.B. Du Bois - Jason Forrest (McKinsey & Co\, New York) DTSTART;VALUE=DATE-TIME:20201006T153000 DTEND;VALUE=DATE-TIME:20201006T170000 UID:https://talks.ox.ac.uk/talks/id/e25a8fac-13bd-4cd6-8d4c-e806c90daf62/ DESCRIPTION:At the 1900 Paris Exposition\, an all African-American team le ad by scholar and activist W.E.B. Du Bois sought to challenge and recontex tualize the perception of African-Americans at the dawn of the 20th-centur y. In less than 5 months\, his team conducted sociological research and ha nd-made more than 60 large data visualization posters for a massive Europe an audience which ultimately awarded Du Bois a gold medal for his efforts. While relatively obscure until recently\, the ramification of his landmar k work remains challenging and especially important in light of the Black Lives Matter movement.\nSpeakers:\nJason Forrest (McKinsey & Co\, New York ) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/e25a8fac-13bd-4cd6-8d4c-e806c90daf62/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:(POSTPONED) Exploring the Data Visualizations of W.E.B. D u Bois - Jason Forrest (McKinsey & Co\, New York) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Cluster-Randomised Test Negative Designs: Inference and Applicatio n to Vector Trials to Eliminate Dengue - Professor Nick Jewell (University of California\, Berkeley School of Public Health) DTSTART;VALUE=DATE-TIME:20200528T160000 DTEND;VALUE=DATE-TIME:20200528T170000 UID:https://talks.ox.ac.uk/talks/id/c4e07240-a753-4f80-be7d-674c94cea488/ DESCRIPTION:The successful introduction of the intracellular bacterium Wol bachia into Aedes aegypti mosquitoes enables a practical approach for deng ue prevention through release of Wolbachia-infected mosquitoes. Wolbachia reduces dengue virus replication in the mosquito and\, once established in the mosquito population\, it is possible that this will provide a long-te rm and sustainable approach to reducing or eliminating dengue transmission . A critical next step is to assess the efficacy of Wolbachia deployments in reducing dengue virus transmission in the field. I will describe and discuss the statistical design of a large-scale cluster randomised test-ne gative parallel arm study to measure the efficacy of such interventions. C omparison of permutation inferential approaches to model based methods wil l be described. Extensions to allow for individual covariates\, and altern ate designs such as the stepped wedge approach\, will also be briefly intr oduced. There are also interesting questions regarding interrupted time-se ries methods associated with analysing pilot site data.\nSpeakers:\nProfes sor Nick Jewell (University of California\, Berkeley School of Public Heal th) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/c4e07240-a753-4f80-be7d-674c94cea488/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Cluster-Randomised Test Negative Designs: Inference and A pplication to Vector Trials to Eliminate Dengue - Professor Nick Jewell (U niversity of California\, Berkeley School of Public Health) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Cluster-Randomised Test Negative Designs: Inference and Applicatio n to Vector Trials to Eliminate Dengue - Professor Nick Jewell (University of California\, Berkeley School of Public Health) DTSTART;VALUE=DATE-TIME:20200528T160000 DTEND;VALUE=DATE-TIME:20200528T170000 UID:https://talks.ox.ac.uk/talks/id/c4e07240-a753-4f80-be7d-674c94cea488/ DESCRIPTION:The successful introduction of the intracellular bacterium Wol bachia into Aedes aegypti mosquitoes enables a practical approach for deng ue prevention through release of Wolbachia-infected mosquitoes. Wolbachia reduces dengue virus replication in the mosquito and\, once established in the mosquito population\, it is possible that this will provide a long-te rm and sustainable approach to reducing or eliminating dengue transmission . A critical next step is to assess the efficacy of Wolbachia deployments in reducing dengue virus transmission in the field. I will describe and discuss the statistical design of a large-scale cluster randomised test-ne gative parallel arm study to measure the efficacy of such interventions. C omparison of permutation inferential approaches to model based methods wil l be described. Extensions to allow for individual covariates\, and altern ate designs such as the stepped wedge approach\, will also be briefly intr oduced. There are also interesting questions regarding interrupted time-se ries methods associated with analysing pilot site data.\nSpeakers:\nProfes sor Nick Jewell (University of California\, Berkeley School of Public Heal th) LOCATION:Venue to be announced TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/c4e07240-a753-4f80-be7d-674c94cea488/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Cluster-Randomised Test Negative Designs: Inference and A pplication to Vector Trials to Eliminate Dengue - Professor Nick Jewell (U niversity of California\, Berkeley School of Public Health) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:GDB and the Chemical Space - Prof Jean-Louis Reymond (Dept. of Che mistry & Biochemistry University of Berne) DTSTART;VALUE=DATE-TIME:20200324T090000Z DTEND;VALUE=DATE-TIME:20200324T100000Z UID:https://talks.ox.ac.uk/talks/id/f36d2b5d-74f3-4787-8add-5c0e6967b597/ DESCRIPTION:Chemical space is a concept to organize molecular diversity by postulating that different molecules occupy different regions of a mathem atical space where the position of each molecule is defined by its propert ies. Our aim is to develop methods to explicitly explore chemical space in the area of drug discovery. We have enumerated all possible molecules fol lowing simple rules of chemical stability and synthetic feasibility to for m the Generated DataBases (GDB). Exploring GDB in comparison to known mol ecules reveals that vast areas of chemical space are still entirely unknow n yet are accessible for experimental exploration by straightforward synth etic methods. I will discuss how to visualize chemical space and exemplify the discovery and synthesis of new scaffolds for drug discovery.\nSpeaker s:\nProf Jean-Louis Reymond (Dept. of Chemistry & Biochemistry University of Berne) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/f36d2b5d-74f3-4787-8add-5c0e6967b597/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:GDB and the Chemical Space - Prof Jean-Louis Reymond (Dep t. of Chemistry & Biochemistry University of Berne) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Maths and Stats in Action – Real-time Analysis to Understand the Novel Coronavirus - Professor Christl Donnelly (University of Oxford)\, D r Robin Thompson (Mathematic Institute) DTSTART;VALUE=DATE-TIME:20200131T130000Z DTEND;VALUE=DATE-TIME:20200131T140000Z UID:https://talks.ox.ac.uk/talks/id/ddb03da4-3f17-4660-b75c-4d7da2a2d4ea/ DESCRIPTION:We will provide a whirlwind tour of the quantitative analyses currently underway to understand the transmission and control of the novel coronavirus (2019-nCOV). \n \nhttps://www.imperial.ac.uk/mrc-global-infe ctious-disease-analysis/news--wuhan-coronavirus/\nhttps://www.biorxiv.org/ content/10.1101/2020.01.24.919159v1\n\nSpeakers:\nProfessor Christl Donnel ly (University of Oxford)\, Dr Robin Thompson (Mathematic Institute) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/ddb03da4-3f17-4660-b75c-4d7da2a2d4ea/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Maths and Stats in Action – Real-time Analysis to Under stand the Novel Coronavirus - Professor Christl Donnelly (University of Ox ford)\, Dr Robin Thompson (Mathematic Institute) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Florence Nightingale and the politicians’ pigeon holes: using da ta for the good of society - Prof Deborah Ashby (Imperial College London) DTSTART;VALUE=DATE-TIME:20200318T143000Z DTEND;VALUE=DATE-TIME:20200318T154500Z UID:https://talks.ox.ac.uk/talks/id/12cf0136-459d-4a0f-8aed-40dd44a4ffc2/ DESCRIPTION:Florence Nightingale\, best known as the Lady with the Lamp\, is recognised as a pioneering and passionate statistician. She was also pa ssionate about education\, having argued successfully with her parents to be allowed to study mathematics\, and later nursing\, herself. More wide ly\, she offered opinions on the education of children\, soldiers\, army d octors\, and nurses\, as well as railing against the ‘enforced idleness ’ of women. A particular concern was the lack of statistical literacy am ong politicians. As we celebrate the bicentenary of her birth\, the need f or education in statistical and data skills shows no signs of abating. Wha t advice would Florence Nightingale offer were she here today?\nSpeakers:\ nProf Deborah Ashby (Imperial College London) LOCATION:Mathematical Institute (L1\, Mathematical Institute\, Andrew Wile s Building\, Woodstock Road\, Oxford)\, Woodstock Road OX2 6GG TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/12cf0136-459d-4a0f-8aed-40dd44a4ffc2/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Florence Nightingale and the politicians’ pigeon holes: using data for the good of society - Prof Deborah Ashby (Imperial College London) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Twenty five year risks of breast cancer mortality in 500\,000 wome n - Prof Sarah C. Darby (University of Oxford) DTSTART;VALUE=DATE-TIME:20200123T153000Z DTEND;VALUE=DATE-TIME:20200123T163000Z UID:https://talks.ox.ac.uk/talks/id/da055ff8-3d17-4987-8332-20fbda7f66cf/ DESCRIPTION:Background: Breast cancer mortality rates vary substantially a cross multiple patient and tumour factors. Large-scale data can show the e ffects of these factors. Long term follow-up is needed because breast canc er deaths continue to occur more than 20 years after the original cancer d iagnosis. Sources of data held by Public Health England on patients with c ancer have increased in recent years and now have the potential to be the best source of population-based information on outcomes worldwide.\n\nMate rials: Large-scale population data were collated on all women registered w ith early invasive breast cancer in England during 1993-2016. Data were av ailable on: year of diagnosis\, age at diagnosis\, mode of presentation (s creen-detected or not)\, tumour size\, number of involved lymph nodes\, gr ade\, oestrogen receptor status\, tumour laterality\, deprivation\, geogr aphical region and date and cause of death up to 31st December 2017. These data were reviewed and improved. Population-based estimates of the absolu te 5\, 10\, 15 and 20-year risks of breast cancer mortality were derived f or women in different prognostic groups.\n\nResults/conclusions: Over 50 0\,000 women were diagnosed with early invasive breast cancer during 1993- 2016\, and around 70\,000 of them subsequently died from breast cancer. Mo re than 20 years’ follow-up was available for around 30\,000 women\, and more than 15 years for around 100\,000 women. Breast cancer mortality var ied substantially according to patient factors such as age at presentation and calendar year of diagnosis and tumour factors such as screening statu s\, size\, node positivity and oestrogen receptor status. Breast cancer mo rtality rates from these large-scale population-based data may be used to estimate mortality for patients diagnosed with breast cancer today. They a re relevant to decisions concerning which treatments and follow-up strateg ies to use.\nSpeakers:\nProf Sarah C. Darby (University of Oxford) LOCATION:24-29 St Giles' (Large Lecture Theatre\, Department of Statistics )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/da055ff8-3d17-4987-8332-20fbda7f66cf/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Twenty five year risks of breast cancer mortality in 500\ ,000 women - Prof Sarah C. Darby (University of Oxford) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Sensitivity Analysis in Observational Research: Introducing the E- Value - Professor Tyler Vanderweele (Departments of Epidemiology and Biost atistics at the Harvard T.H. Chan School of Public Health) DTSTART;VALUE=DATE-TIME:20191129T153000Z DTEND;VALUE=DATE-TIME:20191129T163000Z UID:https://talks.ox.ac.uk/talks/id/2559865e-5649-4df7-9082-b2cc0e9bf4b8/ DESCRIPTION:Sensitivity analysis is useful in assessing how robust an asso ciation is to potential unmeasured or uncontrolled confounding. This artic le introduces a new measure called the “E-value\,” which is related to the evidence for causality in observational studies that are potentially subject to confounding. The E-value is defined as the minimum strength of association\, on the risk ratio scale\, that an unmeasured confounder woul d need to have with both the treatment and the outcome to fully explain aw ay a specific treatment–outcome association\, conditional on the measure d covariates. A large E-value implies that considerable unmeasured confoun ding would be needed to explain away an effect estimate. A small E-value i mplies little unmeasured confounding would be needed to explain away an ef fect estimate. The authors propose that in all observational studies inten ded to produce evidence for causality\, the E-value be reported or some ot her sensitivity analysis be used. They suggest calculating the E-value for both the observed association estimate (after adjustments for measured co nfounders) and the limit of the confidence interval closest to the null. I n observational studies\, the E-value provides an important supplement to the p-value. If this were to become standard practice\, the ability of the scientific community to assess evidence from observational studies would improve considerably\, and ultimately\, science would be strengthened. Que stions of interpretation and relations with prior sensitivity analysis tec hniques and Rosenbaum’s design sensitivity will be discussed.\nSpeakers: \nProfessor Tyler Vanderweele (Departments of Epidemiology and Biostatisti cs at the Harvard T.H. Chan School of Public Health) LOCATION:24-29 St Giles' (Large Lecture Theatre (LG.01))\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/2559865e-5649-4df7-9082-b2cc0e9bf4b8/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Sensitivity Analysis in Observational Research: Introduci ng the E-Value - Professor Tyler Vanderweele (Departments of Epidemiology and Biostatistics at the Harvard T.H. Chan School of Public Health) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Scattered thoughts from applied probability: networks\, security queues and prediction tournaments - Prof David Aldous (University of Calif ornia\, Berkeley) DTSTART;VALUE=DATE-TIME:20190614T153000 DTEND;VALUE=DATE-TIME:20190614T163000 UID:https://talks.ox.ac.uk/talks/id/ccd94884-102e-473b-add9-4ca68e84bffa/ DESCRIPTION:A sample of topics that have caught my eye in recent years.\n\ n(a) Prediction tournaments are like sports in that a higher-ranked player will likely score better than a lower-ranked one. But paradoxically\, un der a reasonable model the winner of a tournament is relatively less likel y to be a top-ranked player\, maybe clouding the interpretation of multimi llion dollar IARPA-sponsored projects.\n\n(b) At the back of a long queue at airport security\, you stand still until a wave of motion reaches you\; a non-standard model gives quantitative predictions for such waves.\n\n( c) Spatial networks give rise to intriguing questions. For instance\, sup pose (as a bizarre fantasy) that an eccentric multi-billionaire proposes t o solve traffic congestion in a huge metropolitan region by digging underg round tunnels through which cars can move very fast\; where to dig the tun nels?\n\n(d) As a hard technical question\, epidemic models combine a mode l for a contact network with a model for transmission between contacts\, w ith various parameters including an infectiousness parameter $\\rho$. Int uitively\, there is always a critical value $\\rho_c$ such that\, starting with a sprinkling of $o(n)$ infectives\, there will w.h.p. be a pandemic if $\\rho > \\rho_c$ but w\,h\,p. not if $\\rho < \\rho_c$. This is true in familiar specific models for the contact network\, but can we prove thi s for all contact networks?\n\nSpeakers:\nProf David Aldous (University of California\, Berkeley) LOCATION:24-29 St Giles' (Large Lecture Theatre (LG.01))\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/ccd94884-102e-473b-add9-4ca68e84bffa/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Scattered thoughts from applied probability: networks\, security queues and prediction tournaments - Prof David Aldous (University of California\, Berkeley) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:From evolutionary trees to networks and back again - Mike Steel (U niversity of Canterbury\, New Zealand) DTSTART;VALUE=DATE-TIME:20190614T140000 DTEND;VALUE=DATE-TIME:20190614T150000 UID:https://talks.ox.ac.uk/talks/id/7c15e755-6a7c-4ccc-89b1-c06226f1b169/ DESCRIPTION:Phylogenetic networks can provide a more complete description of evolutionary history than trees\, by allowing reticulate events such as hybridization and lateral gene transfer. In the first part of this talk\ , we explore the question: ‘when is a phylogenetic network merely a tree with additional links between its edges?’ It turns out that the class o f ‘tree-based’ networks can be efficiently characterized. More recent results on this question have followed\, motivated by Dilworth’s theore m (for posets)\, and matching theory in bipartite graphs. This allows for fast algorithms to determine when a network is tree-based and\, if not\, to calculate how ‘close’ to tree-based it is. In the second part of t he talk\, we model lateral gene transfer by a simple stochastic process on trees. By connecting this process to a simple random walk on a graph it i s possible to analyze the extent to which an underlying species tree T can be inferred from sampled gene trees that have undergone random transfers on T. \nSpeakers:\nMike Steel (University of Canterbury\, New Zealand) LOCATION:24-29 St Giles' (Large Lecture Theatre (LG.01))\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/7c15e755-6a7c-4ccc-89b1-c06226f1b169/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:From evolutionary trees to networks and back again - Mike Steel (University of Canterbury\, New Zealand) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Individualizing Healthcare with Machine Learning - Professor Suchi Saria (Department of Computer Science\, Johns Hopkins University\, USA) DTSTART;VALUE=DATE-TIME:20181130T153000Z DTEND;VALUE=DATE-TIME:20181130T163000Z UID:https://talks.ox.ac.uk/talks/id/93f4d880-b436-4feb-82c8-754026cb67de/ DESCRIPTION:Healthcare is rapidly becoming a data-intensive discipline\, d riven by increasing digitization of health data\, novel measurement techno logies\, and new policy-based incentives. Critical decisions about whom a nd how to treat can be made more precisely by layering an individual’s d ata over that from a population. In my laboratory\, we develop new classes of computational diagnostic and treatment planning tools—tools that te ase out subtle information from “messy” observational datasets\, and p rovide reliable inferences given detailed context about the individual pat ient. I will give example disease areas where such tools are already begin ning to show translational impact. In context\, I will describe challenges associated with learning models from these data and new techniques that l everage probabilistic methods and counterfactual reasoning for tackling th e aforementioned challenges. \nSpeakers:\nProfessor Suchi Saria (Departmen t of Computer Science\, Johns Hopkins University\, USA) LOCATION:24-29 St Giles' (Department of Statistics\, Large Lecture Theatre )\, 24-29 St Giles' OX1 3LB TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/93f4d880-b436-4feb-82c8-754026cb67de/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Individualizing Healthcare with Machine Learning - Profes sor Suchi Saria (Department of Computer Science\, Johns Hopkins University \, USA) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:The Sample Complexity of Multi-Reference Alignment - Philippe Rigo llet\, (MIT Mathematics\, USA) DTSTART;VALUE=DATE-TIME:20180622T153000 DTEND;VALUE=DATE-TIME:20180622T163000 UID:https://talks.ox.ac.uk/talks/id/5ecfa4a1-aefb-4b10-8b9e-7d057aa6fe3a/ DESCRIPTION:How should one estimate a signal\, given only access to noisy versions of the signal corrupted by unknown cyclic shifts? This simple pro blem has surprisingly broad applications\, in fields from aircraft radar i maging to structural biology with the ultimate goal of understanding the s ample complexity of Cryo-EM. We describe how this model can be viewed as a multivariate Gaussian mixture model whose centers belong to an orbit of a group of orthogonal transformations. This enables us to derive matching l ower and upper bounds for the optimal rate of statistical estimation for t he underlying signal. These bounds show a striking dependence on the signa l-to-noise ratio of the problem. We also show how a tensor based method of moments can solve the problem efficiently. Based on joint work with Afon so Bandeira (NYU)\, Amelia Perry (MIT)\, Amit Singer (Princeton) and Jonat han Weed (MIT).\nSpeakers:\nPhilippe Rigollet\, (MIT Mathematics\, USA) LOCATION:Large Lecture Theatre TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/5ecfa4a1-aefb-4b10-8b9e-7d057aa6fe3a/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:The Sample Complexity of Multi-Reference Alignment - Phil ippe Rigollet\, (MIT Mathematics\, USA) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Patterns and surprises in rich but noisy network data - Mark Newma n (University of Michigan\, USA) DTSTART;VALUE=DATE-TIME:20180518T153000 DTEND;VALUE=DATE-TIME:20180518T163000 UID:https://talks.ox.ac.uk/talks/id/fea32dc1-e835-45a1-b622-3b8dfe1f7d6b/ DESCRIPTION:tbc\nSpeakers:\nMark Newman (University of Michigan\, USA) LOCATION:Large Lecture Theatre TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/fea32dc1-e835-45a1-b622-3b8dfe1f7d6b/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Patterns and surprises in rich but noisy network data - M ark Newman (University of Michigan\, USA) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY:Algorithms and Algorithmic Obstacles in High-Dimensional Regressio n - David Gamarnik (MIT Sloan School of Management\, USA) DTSTART;VALUE=DATE-TIME:20180427T153000 DTEND;VALUE=DATE-TIME:20180427T163000 UID:https://talks.ox.ac.uk/talks/id/cf2bfe86-9395-49d1-a6bc-e8ee10e644be/ DESCRIPTION:Many optimization problems arising in studying of random struc tures exhibit an apparent gap between the optimal values which can be esti mated by non-constructive means\, and the best values achievable by fast ( polynomial time) algorithms. Through a combined effort of mathematicians\, computer scientists and statistical physicists\, it became apparent that a potential and in some cases a provable obstruction for designing algorit hms bridging this gap is a phase transition in the geometry of nearly opti mal solutions\, in particular the presence of a certain Overlap Gap Proper ty (OGP).\n\nIn this talk we discuss this property in the context of spars e high dimensional linear regression problem with Gaussian design. We show that\, on the one hand\, in the sampling regime where the known fast meth ods for this problem are effective\, the space of solutions exhibits a mon otonicity with respect to the proximity to the ground truth regression vec tor and no local optimums exist apart from the ground truth. On the other hand\, once the sampling number is asymptotically in the regime where the known methods fail\, we show that the monotonicity is lost\, and the model exhibits an OGP. In the context of the regression problem this means ever y solution exhibiting a small mean squared error is either fairly close to the ground truth or is very far from it\, with no middle ground.\nSpeaker s:\nDavid Gamarnik (MIT Sloan School of Management\, USA) LOCATION:Large Lecture Theatre TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/cf2bfe86-9395-49d1-a6bc-e8ee10e644be/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk:Algorithms and Algorithmic Obstacles in High-Dimensional Regression - David Gamarnik (MIT Sloan School of Management\, USA) TRIGGER:-PT1H END:VALARM END:VEVENT BEGIN:VEVENT SUMMARY: (Why) do functional sites induce long-range evolutionary constrai nts in enzymes? - Julian Echave (Universidad Nacional de San Martín\, Bue nos Aires\, Argentina) DTSTART;VALUE=DATE-TIME:20180420T153000 DTEND;VALUE=DATE-TIME:20180420T163000 UID:https://talks.ox.ac.uk/talks/id/1818d659-b80c-4ec4-8f15-a55ff8877d5e/ DESCRIPTION:Protein evolution can be viewed as a repeated mutation-fixatio n process. At each step\, one amino acid is randomly mutated. The mutant w ill eventually be either discarded or fixed\, replacing the parent. The fi xation probability varies from site to site\, thus different sites evolve at different rates. This variation of rates among sites is due to thermody namic and/or functional constraints.\n\nAlmost all biophysical models of p rotein evolution developed so far consider only selection for stability\, disregarding functional constraints\, which were thought to affect only a handful of residues: the active site and its immediate neighbours. However \, a recent study showed that site-specific evolutionary rates increase ra ther smoothly with increasing distance to the protein’s active residues. Such long-range rate-distance dependence cannot be explained with current stability-based models and is at odds with the localized fast exponential decrease of coupling strength one would expect on physics grounds.\n\nTo understand whether and why functional constraints have long-range effects\ , we need new models of protein evolution that go beyond selection for sta bility and consider protein activity explicitly. Here\, I will describe a model of protein evolution that considers explicitly both stability and ac tivity constraints and I will discuss its predictions. The stability-activ ity model predicts that fitness cost is localized (short-range) yet rate v ariation is delocalized (long-range)\; short-range fitness effects are con sistent with long-range rate effects. Yet\, such long-range coupling is no t universal but range varies among proteins. Such range variation among pr oteins does not depend on intrinsic protein properties but on external fun ctional selection pressure (e.g. the role of an enzyme in the metabolic ne twork).\nSpeakers:\nJulian Echave (Universidad Nacional de San Martín\, B uenos Aires\, Argentina) LOCATION:Large Lecture Theatre TZID:Europe/London URL:https://talks.ox.ac.uk/talks/id/1818d659-b80c-4ec4-8f15-a55ff8877d5e/ BEGIN:VALARM ACTION:display DESCRIPTION:Talk: (Why) do functional sites induce long-range evolutionary constraints in enzymes? - Julian Echave (Universidad Nacional de San Mart ín\, Buenos Aires\, Argentina) TRIGGER:-PT1H END:VALARM END:VEVENT END:VCALENDAR