BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Christophe Biscio (Aalborg University) DTSTART:20220929T131500Z DTEND:20220929T140000Z DTSTAMP:20260404T095849Z UID:gbgstats/1 DESCRIPTION:Title: Asymptotic topological data analysis for point processes\nby C hristophe Biscio (Aalborg University) as part of Gothenburg statistics sem inar\n\nLecture held in MVL14.\n\nAbstract\nTopological Data Analysis has in the past year attracted more attention in various fields such as in mat erial sciences to study the properties of porous material or in statistics to study the asymptotic properties of random objects. However\, topologic al data analysis still appears hard to grasp for many statisticians. \n\nT his talk intends to be an introduction to topological data analysis and th erefore does not require any background in the field. We will present an o verview of the different approaches in topological data analysis and will focus on the persistent homology approach. \nWe will present the framework of this approach and its main mathematical objects. \nFinally\, we come b ack to the land of Probability and will present a central limit theorem fo r the so-called Betti numbers obtained from stationary point processes\, n on-necessarily Poisson.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/1/ END:VEVENT BEGIN:VEVENT SUMMARY:Anders Ståhlberg & Serik Sagitov (Chalmers & University of Gothen burg) DTSTART:20221006T131500Z DTEND:20221006T140000Z DTSTAMP:20260404T095849Z UID:gbgstats/2 DESCRIPTION:Title: Counting molecular identifiers in sequencing using a multitype bra nching process with immigration\nby Anders Ståhlberg & Serik Sagitov (Chalmers & University of Gothenburg) as part of Gothenburg statistics sem inar\n\nLecture held in MVL14.\n\nAbstract\nDetection of extremely rare va riant alleles\, such as tumour DNA\, within a complex mixture of DNA molec ules is experimentally challenging due to sequencing errors. Barcoding of target DNA molecules in library construction for next-generation sequencin g provides a way to identify and bioinformatically remove polymerase induc ed errors. During the barcoding procedure involving $t$ consecutive PCR cy cles\, the DNA molecules become barcoded by unique molecular identifiers ( UMI). Different library construction protocols utilise different values of $t$. The effect of a larger $t$ and imperfect PCR amplifications is poorl y described. \n\nThis paper proposes a branching process with growing immi gration as a model describing the random outcome of $t$ cycles of PCR ba rcoding. Our model discriminates between five different amplification rate s $r_1$\, $r_2$\, $r_3$\, $r_4$\, $r$ for different types of molecules ass ociated with the PCR barcoding procedure. We study this model by focussing on $C_t$\, the number of clusters of molecules sharing the same \nUMI\, as well as $C_t(m)$\, the number of UMI clusters of size $m$. Our main fi nding is a remarkable asymptotic pattern valid for moderately large $t$. I t turns out that \n$E(C_t(m))/E(C_t)\\approx 2^{-m}$ for $m=1\,2\,\\ldots$ \, regardless of the underlying parameters $(r_1\,r_2\,r_3\,r_4\,r)$. The knowledge of the quantities $C_t$ and $C_t(m)$ as functions of the experim ental parameters $t$ and $(r_1\,r_2\,r_3\,r_4\,r)$ will help the users to draw more adequate conclusions from the outcomes of different sequencing p rotocols.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/2/ END:VEVENT BEGIN:VEVENT SUMMARY:Peter Guttorp (University of Washington/Norwegian computing center ) DTSTART:20221027T131500Z DTEND:20221027T140000Z DTSTAMP:20260404T095849Z UID:gbgstats/3 DESCRIPTION:Title: Comparing recent climate models to data\nby Peter Guttorp (Uni versity of Washington/Norwegian computing center) as part of Gothenburg st atistics seminar\n\nLecture held in MVL14.\n\nAbstract\nThe latest climate model intercomparison project (CMIP6) was the basis for the sixth assessm ent report of the Intergovernmental Panel on Climate Change. The design of CMIP6 included climate runs with historical forcings\, meant to be compar able to observational data. We will focus on global annual mean temperatur e\, a common (but not particularly sensitive) measure of climate change. U sing four observational products provided with uncertainty assessments\, w e combine these into a single series. In doing so\, we estimate a smooth t rend and a residual spectral density function\, with attendant simultaneou s confidence bands. Using the same kind of decomposition of 318 climate mo del runs from 58 models in the historical CMIP6 experiment\, we see how we ll the model runs agree with the data. We also compare the warming between 1880-1899 and 1995-2014. This is joint work with Peter Craigmile of the O hio State University.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/3/ END:VEVENT BEGIN:VEVENT SUMMARY:Olle Häggström (Chalmers University of Technology & University o f Gothenburg) DTSTART:20221013T131500Z DTEND:20221013T140000Z DTSTAMP:20260404T095849Z UID:gbgstats/4 DESCRIPTION:Title: Anthropic reasoning and the hinge of history hypothesis\nby Ol le Häggström (Chalmers University of Technology & University of Gothenbu rg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\n Abstract\nAmong researchers and scholars working on existential risk and t he long-term future of humanity\, it has become increasingly common to spe ak of our present time as uniquely pivotal for the long-term future - a no tion that has become known under the term Hinge of History (HoH). Recently \, attempts have been made to formalize this concept and work out whether we really do live during the HoH. This involves not only a Bayesian analys is but also controversial ideas in anthropic reasoning. I will review and critique this work and arrive at a nuanced answer to the HoH question and its practical ramifications.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/4/ END:VEVENT BEGIN:VEVENT SUMMARY:Moritz Schauer (Chalmers University of Technology & University of Gothenburg) DTSTART:20221020T131500Z DTEND:20221020T140000Z DTSTAMP:20260404T095849Z UID:gbgstats/5 DESCRIPTION:Title: Automatic differentiation of programs with discrete randomness \nby Moritz Schauer (Chalmers University of Technology & University of Got henburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14 .\n\nAbstract\nAutomatic differentiation (AD)\, a technique for constructi ng new programs which compute the derivative of an original program\, has become ubiquitous throughout scientific computing and deep learning due to the improved performance afforded by gradient-based optimization. However \, AD systems have been restricted to the subset of programs that have a c ontinuous dependence on parameters. Programs that have discrete stochastic behaviors governed by distribution parameters\, such as flipping a coin w ith probability p of being heads\, pose a challenge to these systems becau se the connection between the result (heads vs tails) and the parameters ( p) is fundamentally discrete. In this paper we develop a new reparameteriz ation-based methodology that allows for generating programs whose expectat ion is the derivative of the expectation of the original program. We showc ase how this method gives an unbiased and low-variance estimator which is as automated as traditional AD mechanisms. We demonstrate unbiased forward -mode AD of discrete-time Markov chains\, agent-based models such as Conwa y's Game of Life\, and unbiased reverse-mode AD of a particle filter. Our code is available at https://github.com/gaurav-arya/StochasticAD.jl\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/5/ END:VEVENT BEGIN:VEVENT SUMMARY:Johan Jonasson (Chalmers University and University of Gothenburg) DTSTART:20230119T141600Z DTEND:20230119T150000Z DTSTAMP:20260404T095849Z UID:gbgstats/6 DESCRIPTION:Title: Noise sensitivity/stability for deep Boolean neural nets\nby J ohan Jonasson (Chalmers University and University of Gothenburg) as part o f Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nA w ell-known and ubiquitous property of neural net classifiers is that they c an be fooled into misclassifying some objects by changing the input in tin y ways that are indistinguishable for the human eye. These changes can be adversarial\, but sometimes they can be just random noise. This makes it i nteresting to ask if this property is something that almost all neural net s have and\, when they do\, why that is. There are good heuristic explanat ions\, but to prove mathematically rigorous results seems very difficult i n general. Here we prove some first results on various toy models. We trea t our questions within the framework of the established field of noise sen sitivity/stability. What we prove can roughly be stated as:\n \n\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/6/ END:VEVENT BEGIN:VEVENT SUMMARY:David Widmann (Uppsala University) DTSTART:20221124T141500Z DTEND:20221124T150000Z DTSTAMP:20260404T095849Z UID:gbgstats/7 DESCRIPTION:Title: Calibration of probabilistic predictive models\nby David Widma nn (Uppsala University) as part of Gothenburg statistics seminar\n\nLectur e held in MVL14.\n\nAbstract\nMost supervised machine learning tasks are s ubject to irreducible prediction errors. Probabilistic predictive models a ddress this limitation by providing probability distributions that represe nt a belief over plausible targets\, rather than point estimates. Such mod els can be a valuable tool in decision-making under uncertainty\, provided that the model output is meaningful and interpretable. Calibrated models guarantee that the probabilistic predictions are neither over- nor under-c onfident. In the machine learning literature\, different measures and stat istical tests have been proposed and studied for evaluating the calibratio n of classification models. For regression problems\, however\, research h as been focused on a weaker condition of calibration based on predicted qu antiles for real-valued targets. In this paper\, we propose the first fram ework that unifies calibration evaluation and tests for general probabilis tic predictive models. It applies to any such model\, including classifica tion and regression models of arbitrary dimension. Furthermore\, the frame work generalizes existing measures and provides a more intuitive reformula tion of a recently proposed framework for calibration in multi-class class ification. In particular\, we reformulate and generalize the kernel calibr ation error\, its estimators\, and hypothesis tests using scalar-valued ke rnels\, and evaluate the calibration of real-valued regression problems.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/7/ END:VEVENT BEGIN:VEVENT SUMMARY:Karin Hårding and Daire Carroll (University of Gothenburg) DTSTART:20221208T141500Z DTEND:20221208T150000Z DTSTAMP:20260404T095849Z UID:gbgstats/8 DESCRIPTION:Title: Population dynamics and ecology of seal populations\, empirical da ta and the search for theory to help our understanding. Stochastic growth models\, image analysis\, spatial distribution and telemetry data on migra tions\nby Karin Hårding and Daire Carroll (University of Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbst ract\nThis talk is about how statistical and mathematical methods are help ful when we try to understand processes in wildlife populations. The Europ ean harbour seal (Sw: knubbsälen) has been studied carefully for 40 years and the long time series allows analysis of how population growth is regu lated. Recently the population growth has declined and we visited the col onies to try to document in detail what is going on in order to give bette r advise to managers. We develop new methods for estimating body size from drones and for counting seals from photos with machine learning algorithm s. We apply stochastic population growth models\, dynamic energy budget mo dels\, and we discuss what is density dependence in age structured populat ions in a variable environment. We are also interested in new collaboratio ns and feed back and look forward to interesting discussions on ways forwa rd. Welcome! Karin and Daire\n\nKarin Harding is professor in animal ecolo gy with a focus on marine mammals at GU and Daire Carroll is postdoctoral researcher and develops new digital tools for wildlife ecology.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/8/ END:VEVENT BEGIN:VEVENT SUMMARY:Khalil Shafie Holighi (University of Northern Colorado) DTSTART:20221201T141500Z DTEND:20221201T150000Z DTSTAMP:20260404T095849Z UID:gbgstats/9 DESCRIPTION:Title: A test for multiple signal detection from noisy images\nby Kha lil Shafie Holighi (University of Northern Colorado) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nGaussian random field theory has been extensively used to model the brain images.\nIn th is work\, I use the reproducing kernel Hilbert space (RKHS) machinery to derive the likelihood ratio test statistic for activation signal detecti on in functional magnetic resonance imaging. The models considered hav e the form of smoothed version of signal plus a white noise which inclu de scale and rotation space random fields with one or more signals as special cases.\n\nKhalil Shafie Holighi is professor of statistics at Univ ersity of Northern Colorado.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/9/ END:VEVENT BEGIN:VEVENT SUMMARY:Peter Guttorp (University of Washington/Norwegian computing center ) DTSTART:20221215T141500Z DTEND:20221215T150000Z DTSTAMP:20260404T095849Z UID:gbgstats/10 DESCRIPTION:Title: Vadan och varthän?\nby Peter Guttorp (University of Washingt on/Norwegian computing center) as part of Gothenburg statistics seminar\n\ nLecture held in MVL14.\n\nAbstract\nSverige har haft professurer is stati stik (och statskunskap) sedan början av 1900-talet. En av dessa var Pontu s Fahlbeck i Lund\, som förespråkade enbart en samhällsvetenskaplig inr iktning på statistiken. Lundaastronomen Carl Charlier sysslade med stella r statistik\, och ansåg att statistikvetenskapen var användbar i naturve tenskap lika väl som samhällskunskap. Vi berättar hur denna konflikt re sulterade i institutioner både i statistik och matematisk statistik. Vi f öreslår att statistikvetenskapen\, vare sig den kallas statistik eller m atematisk statistik\, bör hamna i en egen institution.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/10/ END:VEVENT BEGIN:VEVENT SUMMARY:Alan Gelfand (Department of Statistical Science\, Duke University) DTSTART:20230316T121500Z DTEND:20230316T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/11 DESCRIPTION:Title: Three Spatial Data Fusion Vignettes\nby Alan Gelfand (Departm ent of Statistical Science\, Duke University) as part of Gothenburg statis tics seminar\n\nLecture held in MVL14.\n\nAbstract\nWith increased collect ion of spatial (and spatio-temporal) datasets\, we often find multiple sou rces that are capable of informing about features of a process of interest . Through suitable fusion of the data sources\, we can learn at least as m uch about the process features of interest than from any individual source . For three different illustrative ecological/environmental applications\ , this talk will propose suitable coherent stochastic modeling to implemen t a fusion of these sources. We focus exclusively on approaches that arise through generative hierarchical modeling\; the specification could produc e the data sources that have been observed. Such modeling enables full in ference both with regard to estimation and prediction\, with implicit inco rporation of uncertainty. \n\nWe consider the general setting of points a nd marks\, modeled as $[points][marks|points]$\, points in $\\mathcal{D}$\ , marks in $\\mathcal{Y}$. The process can model the points themselves\, the marks themselves (ignoring any randomness in the points)\, or the poin ts and marks jointly. This results in four data types: (i) a point patter n\, $\\mathcal{S}= (\\textbf{s}_{1}\, \\textbf{s}_{2}\,\\ldots\,\\textbf{s }_{n})$\, (ii) a vector of counts for sets\, $\\{N(B_{k})\, k=1\,2\,\\ldot s\,K\\}$\, (iii) a vector of observations at points\, $\\{Y(\\textbf{s}_{i })\,i=1\,2\,\\ldots\,n\\}$\, (iv) a vector of averages for sets\, $\\{Y(B_ {1})\, Y(B_{2})\,\\ldots\,Y(B_{k})\\}$. We illustrate with two data sourc es\; each can be any one of the four data types. Regardless of how the da ta are observed\, we imagine the process operates at point level. Further\ , we imagine a stochastic process over $\\mathcal{D}$ which links the two data sources.\n\nThe first vignette considers presence/absence data over $ \\mathcal{D}$ with one dataset being presence/absence of a species collect ed at a set of chosen locations. The other data source is in the form of museum/citizen science data\, recording random locations where the species was observed. The goal is to better understand the probability of presen ce surface over $\\mathcal{D}$. The second vignette considers zooplankton abundance data gathered through two different $\\it{towing}$ mechanisms. One mechanism is calibrated while the other is not. The goal is to better understand zooplankton abundance over $\\mathcal{D}$. The third\, and mo st challenging vignette seeks to learn about whale abundance. Here\, the two sources are aerial distance sampling data for whale sightings and pass ive acoustic monitoring data (using monitors on the ocean floor) for whale calls. \n\nThis is joint work with Shin Shirota\, Jorge Castillo-Mateo\, Erin Schliep\, and Rob Schick.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/11/ END:VEVENT BEGIN:VEVENT SUMMARY:Martin Voigt Vejling (Aalborg University) DTSTART:20230216T121500Z DTEND:20230216T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/12 DESCRIPTION:Title: Applications of point process models to wireless communication sy stems\nby Martin Voigt Vejling (Aalborg University) as part of Gothenb urg statistics seminar\n\nLecture held in MVL15.\n\nAbstract\nWireless com munications is a field in engineering that is attracting a lot of attentio n and one of the envisioned enablers of future communication systems is st atistical learning. Interestingly\, many scenarios considered in wireless communications is naturally modelled by point processes. This has historic ally been studied with stochastic geometry models for wireless networks\, however\, some areas of applications remain underexplored.\n\nIn this talk \, I will introduce the audience to what wireless communication is\, discu ssing the basic concepts\, models\, and challenges. This includes discussi ng the typically used parametric channel model and how harmonic analysis\, estimation theory\, and compressive sensing are central mathematical topi cs applied in practice. Then\, I will motivate the use of point process mo dels within wireless communications by giving a general modelling perspect ive. Finally\, I will discuss the primary focus of my research in radio fr equency sensing powered by statistical learning for point processes.\n\nTh e speaker is a PhD student shared between the Department of Electronic Sys tems and the Department of Mathematical Sciences at Aalborg University. Th e talk is intended for researchers in spatial statistics interested in app lications to topics in wireless communications.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/12/ END:VEVENT BEGIN:VEVENT SUMMARY:Torgny Lindvall (Chalmers University of Technology & University of Gothenburg) DTSTART:20230202T121500Z DTEND:20230202T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/13 DESCRIPTION:Title: On coupling of renewal processes and random walks\nby Torgny Lindvall (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL15.\n\nAbstra ct\nWe use an Ornstein coupling for another proof of Blackwell's renewal t heorem\, and a Mineka coupling to establish a 0-2 law for random walks.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/13/ END:VEVENT BEGIN:VEVENT SUMMARY:Mike Pereira (Mines Paris - PSL University) DTSTART:20230223T121500Z DTEND:20230223T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/15 DESCRIPTION:Title: Gaussian fields on Riemannian manifolds: Application to Geostatis tics.\nby Mike Pereira (Mines Paris - PSL University) as part of Gothe nburg statistics seminar\n\nLecture held in MVL15.\n\nAbstract\nMany appli cations in spatial statistics require data to be modeled by Gaussian proce sses on non-Euclidean domains\, or with non-stationary properties. Using such models generally comes at the price of a drastic increase in operatio nal costs (computational and storage-wise)\, rendering them hard to apply to large datasets. In this talk\, we propose a solution to this problem\, which relies on the definition of a class of random fields on Riemannian m anifolds. These fields extend ongoing work that has been done to leverage a characterization of the random fields classically used in Geostatistics as solutions of stochastic partial differential equations. The discretizat ion of these generalized random fields\, undertaken using a finite element approach\, then provides an explicit characterization that is leveraged t o solve the scalability problem. Indeed\, matrix-free algorithms\, in the sense that they do not require to build and store any covariance (or preci sion) matrix\, are derived to tackle for instance the simulation of large Gaussian fields with given covariance properties\, even in the non-station ary setting.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/15/ END:VEVENT BEGIN:VEVENT SUMMARY:Konstantinos Konstantinou (Chalmers University of Technology & Uni versity of Gothenburg) DTSTART:20230309T121500Z DTEND:20230309T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/16 DESCRIPTION:Title: Global tests for quantile regression with applications in modelin g distributions.\nby Konstantinos Konstantinou (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistics s eminar\n\nLecture held in MVL15.\n\nAbstract\nIn this talk\, I will give a n introduction to global tests for quantile regression\, i.e.\, statistica l tests allowing for simultaneous inference of the quantile regression pro cess\, with graphical interpretation. The proposed global quantile regress ion tests can determine not only if there is a difference\, but it can als o determine for which quantiles the difference is significant on the globa l significance level. The case where the effect of a factor (e.g.\, a cate gorical factor giving the group) on the distribution functions is of inter est but confounded with other factors affecting the distributions is studi ed. An extensive simulation study is conducted to compare the global quant ile regression tests with classical graphical tests based on the Kolmogoro v-Smirnov test statistic. This is a joint work with Tomáš Mrkvička\, Ma ri Myllymäki and Mikko Kuronen.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/16/ END:VEVENT BEGIN:VEVENT SUMMARY:Finn Lindgren (University of Edinburgh) DTSTART:20230314T121500Z DTEND:20230314T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/17 DESCRIPTION:Title: Stochastic adventures in space and time\nby Finn Lindgren (Un iversity of Edinburgh) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nThe standard geostatistics toolbox includes m ethods for modelling\nspatial dependence between georeferenced observation s\, as well as\nmethods for modelling the occurrence of random points. Th e core\nmodel building blocks are often some form of Gaussian random field s.\n\nThe easiest approach to constructing space-time models is by taking\ nthe product between a spatial covariance kernel and a temporal\ncovarianc e kernel. These are called covariance separable models. An\nalternative th at may better capture the spatio-temporal dynamics is to\ntake inspiration for physics motivated partial differential equations\nsuch as the heat eq uation\, which leads to non-separable models.\nNon-separable models are in general more computationally expensive\,\nbut one can sometimes use the m odel structure to retain a lot of the\nsimplicity of separable models\, fo r example allowing these models to\nbe used as components of larger hierar chical generalised additive\nmodels. For point process observations\, such as observations of a\nmoving animal\, the temporal dynamics poses an addi tional challenge.\n\nI will discuss some of these aspects\, including a ba sic construction\nof non-separable space-time models\, as well as an appli cation of the\nINLA/inlabru framework to estimate the parameters of a dyna mical\nanimal movement model by rephrasing it as a point process model\, w ith\na parametric movement kernel\, and a random field as an unknown\n"res ource selection function".\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/17/ END:VEVENT BEGIN:VEVENT SUMMARY:Nicoletta D’Angelo (University of Palermo) DTSTART:20230404T111500Z DTEND:20230404T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/18 DESCRIPTION:Title: Self-exciting point process modelling of crimes on linear network s\nby Nicoletta D’Angelo (University of Palermo) as part of Gothenbu rg statistics seminar\n\nLecture held in MVL15.\n\nAbstract\nAlthough ther e are recent developments in analysing first and second-order characterist ics of point processes on networks\, there are very few attempts to introd uce models for network data.\nMotivated by the analysis of crime data in B ucaramanga (Colombia)\, we propose a spatio-temporal Hawkes point process model adapted to events living on linear networks. We first consider a no n-parametric modelling strategy\, for which we follow a non-parametric est imation of both the background and the triggering components. Then we cons ider a semi-parametric version\, including a parametric estimation of the background based on covariates. Our network model outperforms a planar ver sion\, improving the fitting of the self-exciting point process model\, an d can be easily adapted to multi-type processes.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/18/ END:VEVENT BEGIN:VEVENT SUMMARY:Petar Jovanovski (Chalmers University of Technology & University o f Gothenburg) DTSTART:20230420T111500Z DTEND:20230420T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/23 DESCRIPTION:Title: Approximate Bayesian Computation with Backward Simulation for Dis cretely Observed Diffusions\nby Petar Jovanovski (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nStochastic differential eq uations (SDE) are employed in many areas of science as a powerful tool for modelling processes that are subject to random fluctuations. Bayesian inf erence for a large class of SDEs is challenging due to the analytic intrac tability of the likelihood function. Nevertheless\, forward simulation via numerical methods is straightforward\, motivating the use of approximate Bayesian computation (ABC). We propose a simulation scheme for SDE models that is based on processing the observation in both the forward and backwa rd direction\, effectively utilizing the information provided by the obser ved data. This leads to the simulation of sample paths that are consistent with the observations\, thereby increasing the ABC acceptance rate. We ad ditionally leverage partial exchangeability of Markov processes and employ invariant neural networks to learn the summary statistics that are needed in ABC. These are sequentially learned by exploiting a sequential Monte C arlo ABC sampler\, which provides new training data at each iteration. The refore\, our novel contribution is a learning tool for SDE model parameter s while simultaneously learning the summary statistics. Using synthetic da ta generated from the Chan-Karaolyi-Longstaff-Sanders SDE family\, we show that our approach accelerates inference considerably\, compared to standa rd (forward-only) methods\, while preserving inference accuracy.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/23/ END:VEVENT BEGIN:VEVENT SUMMARY:David Bolin (King Abdullah University of Science and Technology) DTSTART:20230427T111500Z DTEND:20230427T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/24 DESCRIPTION:Title: Gaussian Whittle-Matérn fields on metric graphs\nby David Bo lin (King Abdullah University of Science and Technology) as part of Gothen burg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nWe define a new class of Gaussian processes on compact metric graphs such as street o r river networks. The proposed models\, the Whittle-Matérn fields\, are d efined via a fractional stochastic partial differential equation on the co mpact metric graph and are a natural extension of Gaussian fields with Mat érn covariance functions on Euclidean domains to the non-Euclidean metric graph setting. Existence of the processes\, as well as their sample path regularity properties are derived. The model class in particular contains differentiable Gaussian processes. To the best of our knowledge\, this is the first construction of a valid differentiable Gaussian field on general compact metric graphs.\nWe then focus on a model subclass which we show c ontains processes with Markov properties. For this case\, we show how to e valuate finite dimensional distributions of the process exactly and comput ationally efficiently. This facilitates using the proposed models for stat istical inference without the need for any approximations. Finally\, we de rive some of the main statistical properties of the model class\, such as consistency of maximum likelihood estimators of model parameters and asymp totic optimality properties of linear prediction based on the model with m isspecified parameters. \nThe usage of the model class is illustrated thro ugh an application to traffic data.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/24/ END:VEVENT BEGIN:VEVENT SUMMARY:Moa Johansson (Chalmers University of Technology) DTSTART:20230504T111500Z DTEND:20230504T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/25 DESCRIPTION:Title: Machine Learning Methods for texts from Political Science\nby Moa Johansson (Chalmers University of Technology) as part of Gothenburg s tatistics seminar\n\nLecture held in MVL14.\n\nAbstract\nIn the WASP-HS pr oject "Bias and Methods of AI Technology Studying Political Behavior" we a re investigating and developing machine learning methods to help political scientists study the enormous amounts of text documents that are otherwis e beyond manual analysis\, such as the document repository from the Swedis h Riksdag. This is a collaborative project between Dr. Annika Fredén's gr oup in the Political Science department at Lund University\, and the group of Dr. Moa Johansson at Computer Science at Chalmers.\n\nI will give an o verview of some of the work so far on how we are trying to highlight diffe rences in language use between parties in the Swedish Riksdag. The first p aper is about comparing word embeddings trained on texts from different pa rties. The second concerns explainability of text classification: if a mac hine learning algorithm can classify text as belonging to one party or ano ther\, it is useful for a social scientist to know what such a classificat ion is based on. We have started to develop a new method for class explain ability for text for this purpose. This is going work with PhD student Den itsa Saynova\, and post-doc Bastiaan Bruinsma.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/25/ END:VEVENT BEGIN:VEVENT SUMMARY:Oskar Allerbo (Chalmers University of Technology & University of G othenburg) DTSTART:20230511T111500Z DTEND:20230511T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/26 DESCRIPTION:Title: Solving Kernel Ridge Regression with Gradient Descent\nby Osk ar Allerbo (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbst ract\nWe present an equivalent formulation for the objective function of k ernel ridge regression (KRR)\, that opens up for studying KRR from the per spective of gradient descent. Utilizing gradient descent with infinitesima l step size\, allows us to formulate a new regularization for kernel regre ssion through early stopping.\n\nThe gradient descent formulation of K RR allows us expand to a time dependent stationary kernel\, where we decre ase the bandwidth to zero during training. This circumvents the need of hy per parameter selection. Furthermore\, we are able to achieve both zer o training error and a double descent behavior\, phenomena that do not occ ur for KRR with constant bandwidth\, but are known to appear for neural ne tworks.\n\nThe new formulation of KRR also enables us to explore other pen alties than the ridge penalty. Specifically\, we explore the $\\ell_1$ and $\\ell_\\infty$ penalties and show that these correspond to two flavors o f gradient descent\, thus alleviating the need of computationally heavy pr oximal gradient descent algorithms. We show theoretically and empirically how these formulations correspond to signal-driven and robust regression\, respectively.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/26/ END:VEVENT BEGIN:VEVENT SUMMARY:Juan Inda (Chalmers University of Technology & University of Gothe nburg) DTSTART:20230516T111500Z DTEND:20230516T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/27 DESCRIPTION:Title: Confidence-based Prediction of Antibiotic Resistance at the Patie nt-level Using Transformers\nby Juan Inda (Chalmers University of Tech nology & University of Gothenburg) as part of Gothenburg statistics semina r\n\nLecture held in MVL15.\n\nAbstract\nRapid and accurate diagnostics of bacterial infections are necessary for efficient treatment of antibiotic- resistant pathogens. Cultivation-based methods\, such as antibiotic suscep tibility testing (AST)\, are slow\, resource-demanding\, and can fail to p roduce results before the treatment needs to start. This increases patient risks and antibiotic overprescription. Here\, we present a deep-learning method that uses transformers to merge patient data with available AST res ults to predict antibiotic susceptibilities that have not been measured. T he method is combined with conformal prediction (CP) to enable the estimat ion of uncertainty at the patient-level. After training on three million A ST results from thirty European countries\, the method made accurate predi ctions for most antibiotics while controlling the error rates\, even when limited diagnostic information was available. We conclude that transformer s and CP enables confidence-based decision support for bacterial infection s and\, thereby\, offer new means to meet the growing burden of antibiotic resistance.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/27/ END:VEVENT BEGIN:VEVENT SUMMARY:Frank Miller (Linköping University and Stockholm University) DTSTART:20230525T111500Z DTEND:20230525T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/28 DESCRIPTION:Title: Parallel optimal pretesting of mixed-format questions for achieve ment tests\nby Frank Miller (Linköping University and Stockholm Unive rsity) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\ n\nAbstract\nWhen large achievement tests like national tests in school or admission tests for university are conducted regularly\, the test-questio ns need to be pretested before being used in the test. Methods for assigni ng pretesting questions to examinees in an optimal way based on their abil ity have been developed. Most of these methods are intended for a situatio n where examinees arrive sequentially for being assigned to pretesting que stions. However\, several pretests (e.g.\, for national tests in Swedish s chools or for högskoleprovet) are conducted in a way where all or many ex aminees conduct the test in parallel. In this talk\, we develop an optimal design for such parallel pretest setups which can be implemented in real scenarios. In many real test situations\, questions are of mixed format an d our optimal design method can handle that. We discuss first the optimal designs for the 2-parameter logistic\, the 3-parameter logistic\, and the generalized partial credit model. Then\, we consider the case of mixed-for mat tests where all these models are used to fit the data. The method we p ropose can also take different expected solve times into consideration. We investigate the efficiency gain of the method. Our investigations show th at the proposed method is able to increase the efficiency of pretests cons iderably. The described method has been used for the Swedish national test s in mathematics. \n\nThis is a joint work with Ellinor Fackle-Fornius\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/28/ END:VEVENT BEGIN:VEVENT SUMMARY:Lotta Eriksson (Chalmers University of Technology & University of Gothenburg) DTSTART:20230601T111500Z DTEND:20230601T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/29 DESCRIPTION:Title: A multitype Galton-Watson model of biological aging\nby Lotta Eriksson (Chalmers University of Technology & University of Gothenburg) a s part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstr act\nThe progressive accumulation of damaged proteins within the cellular structure is a key factor in the aging process of yeast. By considering th e quantity of damaged proteins as a measure of the cell's biological age\, we explore an individual-based stochastic population model that incorpora tes rejuvenation events. During cell division\, the mother cell is given t he opportunity rejuvenate\, by transferring the accumulated damage to the daughter cell. This modeling approach allows us to study the dynamics of t he aging process and understand the impact of rejuvenation on the overall population.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/29/ END:VEVENT BEGIN:VEVENT SUMMARY:Felix Held (Chalmers University of Technology & University of Goth enburg) DTSTART:20230608T111500Z DTEND:20230608T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/30 DESCRIPTION:Title: Simultaneous gene clustering and regulatory program reconstructio n reveals insight into the phenotypic plasticity of neural cancers\nby Felix Held (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbs tract\nNervous system cancers contain a large spectrum of transcriptional cell states\, reflecting processes active during normal development\, inju ry response and growth. However\, we lack a good understanding of these st ates' regulation and pharmacological importance. Here\, we describe the in tegrated reconstruction of such cellular regulatory programs and their the rapeutic targets from extensive collections of single-cell RNA sequencing data (scRNA-seq). Our approach called single-cell Regulatory-driven Cluste ring (scRegClust) performs simultaneous gene clustering and regulatory pro gram reconstruction tasks to predict essential kinases and transcription f actors. We formulate an apriori intractable partitioning problem that conn ects gene modules with linear regulator models. A greedy two-step procedur e is constructed to iteratively update gene modules and associated regulat ory programs and find an approximate solution. Penalized regression was us ed to replace a combinatorial selection problem in the construction of reg ulatory programs and predictive modelling was used during gene cluster all ocation. The method is used to identify regulatory programs in tumor cell states from both adult and childhood brain cancers. Further analysis corro borated by experimental results leads to hypothesis generation on an under lying biological mechanism for drug combination therapy in adult glioblast oma.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/30/ END:VEVENT BEGIN:VEVENT SUMMARY:Sara Hamis (Tampere University) DTSTART:20231026T111500Z DTEND:20231026T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/31 DESCRIPTION:Title: Spatial cumulant models for mathematical cancer research\nby Sara Hamis (Tampere University) as part of Gothenburg statistics seminar\n \nLecture held in MVL14.\n\nAbstract\nSpatial cumulant models (SCMs) are s patially resolved population models\, formulated by differential equations \, that describe population dynamics generated by spatio-temporal point pr ocesses (STPPs). Specifically\, SCMs approximate the dynamics of two STPP- generated summary statistics: first-order spatial cumulants (densities)\, and second-order spatial cumulants (spatial covariances). \n\nIn this talk \, I’ll exemplify how SCMs can be used in mathematical oncology by model ling theoretical cancer cell populations comprising interacting subclones. Our results demonstrate that SCMs can capture STPP-generated population d ensity dynamics\, even when mean-field population models (MFPMs) fail to d o so. From both MFPM and SCM equations\, we derive treatment-induced death rates required to achieve non-growing cell populations. When testing thes e treatment strategies in STPP-generated cell populations\, our results de monstrate that SCM-informed strategies outperform MFPM-informed strategies in terms of inhibiting population growths. We thus demonstrate that SCMs provide a new framework in which to study cell-cell interactions and treat ments that take cell-cell interactions into account. \n\nJoint work with: Panu Somervuo\; J. Arvid Ågren\; Dagim S. Tadele\; Juha Kesseli\; Jacob G . Scott\; Matti Nykter\; Philip Gerlee\; Dmitri Finkelshtein\; Otso Ovaska inen.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/31/ END:VEVENT BEGIN:VEVENT SUMMARY:Hrvoje Planinić (University of Zagreb) DTSTART:20230921T091500Z DTEND:20230921T100000Z DTSTAMP:20260404T095849Z UID:gbgstats/32 DESCRIPTION:Title: Extremes of stationary heavy-tailed time series\nby Hrvoje Pl aninić (University of Zagreb) as part of Gothenburg statistics seminar\n\ nLecture held in MVL14.\n\nAbstract\nWe will present a framework for descr ibing the asymptotic behavior of high-level exceedances for stationary (i. e. dependent) time series with heavy-tailed marginal distribution and whos e exceedances occur in clusters\; think of modelling e.g. financial return s or daily rainfall measurements. The main tools are the theory of point p rocesses and the notion of the so-called tail process. The latter allows o ne to fully describe the asymptotic distribution of the extremal clusters using the language of standard Palm theory. We will illustrate the general theory on simple moving average models. If time permits\, we will comment on how this framework can be extended to deal with extremes related to mo dels from stochastic geometry.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/32/ END:VEVENT BEGIN:VEVENT SUMMARY:Devdatt Dubhashi (Chalmers) DTSTART:20230928T111500Z DTEND:20230928T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/33 DESCRIPTION:Title: Bandits: Structured and Constrained\nby Devdatt Dubhashi (Cha lmers) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\ n\nAbstract\nI will give an introduction and invitation to Bandits - a ver y simple\, yet central model of sequential decision making under uncertain ty. After introducing the central concepts and some of the basic algorithm s and results\, I'll describe some recent work from our group on some exte nsions of the basic model which are also useful in applications. Throughou t I'll try to show how the subject has close connections to information th eory\, statistics and optimization.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/33/ END:VEVENT BEGIN:VEVENT SUMMARY:Serik Sagitov (Chalmers University of Technology & university of G othenburg) DTSTART:20231103T121500Z DTEND:20231103T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/35 DESCRIPTION:Title: Theta-positive branching processes in varying environment\nby Serik Sagitov (Chalmers University of Technology & university of Gothenbu rg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\n Abstract\nBranching processes in a varying environment encompass a wide ra nge of stochastic demographic models\, and their complete understanding in terms of limit behavior poses a formidable research challenge. In this pa per\, we conduct a thorough investigation of such processes within a conti nuous-time framework\, assuming that the reproduction law of individuals a dheres to a specific parametric form for the probability generating functi on. Our six clear-cut limit theorems support the notion of recognizing fiv e distinct asymptotical regimes for branching in varying environments: sup ercritical\, asymptotically degenerate\, critical\, strictly subcritical\, and loosely subcritical.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/35/ END:VEVENT BEGIN:VEVENT SUMMARY:Olle Häggström (Chalmers University of Technology & university o f Gothenburg) DTSTART:20231006T111500Z DTEND:20231006T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/36 DESCRIPTION:Title: Playing with fire\nby Olle Häggström (Chalmers University o f Technology & university of Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nAlan Turing speculated in 1 951 about a time point in the future when machines “outstrip our feeble powers” in such a way that we lose our position as the most intelligent species on the planet. Current AI trends suggest that we are rapidly appro aching that time point. This is playing with fire\, because at such a time point our continued wellbeing hinges largely on what the first superintel ligent machines are motivated to do. If their goals and values are aligned with ours\, then a brilliant future awaits us\, while if not\, then most likely it is (in the words of OpenAI’s CEO Sam Altman) “lights out for everyone”. Making this transition go well involves considerable technol ogical and societal challenges.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/36/ END:VEVENT BEGIN:VEVENT SUMMARY:Moritz Schauer (Chalmers University of Technology & University of Gothenburg) DTSTART:20231020T090000Z DTEND:20231020T094500Z DTSTAMP:20260404T095849Z UID:gbgstats/37 DESCRIPTION:Title: Causal structure learning and sampling using Markov Monte Carlo w ith momentum\nby Moritz Schauer (Chalmers University of Technology & U niversity of Gothenburg) as part of Gothenburg statistics seminar\n\n\nAbs tract\nIn the context of inferring a Bayesian network structure from obser vational data\, that is inferring a directed acyclic graph (DAG)\, we devi se a non-reversible continuous-time Markov chain that targets a probabilit y distribution over classes of observationally equivalent (Markov equivale nt) DAGs. The classes are represented as completed partially directed acyc lic graphs (CPDAGs). The non-reversible Markov chain relies on the operato rs used in Chickering’s Greedy Equivalence Search (GES) and is endowed w ith a momentum variable\, which improves mixing significantly as we show e mpirically. The possible target distributions include posterior distributi ons based on a prior and a Markov equivalent likelihood. Joint work with M arcel Wienöbst (Universität zu Lübeck).\n\nThis is a talk in the webina r series of the Cramér society\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/37/ END:VEVENT BEGIN:VEVENT SUMMARY:Frédéric Lavancier (Nantes University\, France) DTSTART:20240119T121500Z DTEND:20240119T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/38 DESCRIPTION:Title: Spatial birth-death-move processes: basic properties and inferenc e\nby Frédéric Lavancier (Nantes University\, France) as part of Got henburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nBirth-de ath-move processes are Markov models for the spatio-temporal dynamics of a system of particles in motion where births and deaths can occur. Natural applications include epidemiology\, individual-based modelling in ecology and spatio-temporal dynamics observed in bio-imaging. We present some of t he basic probabilistic properties of these processes and we consider two i nference problems: First\, the non-parametric estimation of the birth and death intensity functions\; Second\, the parametric estimation of the full dynamics by maximum likelihood. We finally apply our statistical method t o the analysis of a real dataset representing the spatio-temporel dynamics of biomolecules observed in a living cell.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/38/ END:VEVENT BEGIN:VEVENT SUMMARY:Christian Hirsch (Aarhus University) DTSTART:20240301T121500Z DTEND:20240301T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/41 DESCRIPTION:Title: On the topology of higher-order age-dependent random connection m odels\nby Christian Hirsch (Aarhus University) as part of Gothenburg s tatistics seminar\n\nLecture held in MVL14.\n\nAbstract\nPreferential atta chment is a popular mechanism for generating scale-free networks. While it offers a compelling narrative\, the underlying reinforced processes make it difficult to rigorously establish subtle properties. Recently\, age-dep endent random connection models were proposed as an alternative that is ca pable of generating similar networks with a mechanism that is amenable to a more refined analysis. In this talk\, we analyze the asymptotic behavior of higher-order topological characteristics such as higher-order degree d istributions and Betti numbers in large domains. We demonstrate the pract ical application of the theoretical results to real-world datasets by anal yzing scientific collaboration networks based on data from arXiv.This talk is based on joint work with Péter Juhász\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/41/ END:VEVENT BEGIN:VEVENT SUMMARY:Pierre Nyquist (Chalmers University of Technology & University of Gothenburg) DTSTART:20240221T121500Z DTEND:20240221T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/42 DESCRIPTION:Title: Large deviations for Markov chain Monte Carlo methods: the surpri singly curious case of Metropolis-Hastings.\nby Pierre Nyquist (Chalme rs University of Technology & University of Gothenburg) as part of Gothenb urg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nMarkov chain Monte Carlo (MCMC) methods have become the workhorse for numerical comput ations in a range of scientific disciplines\, e.g.\, computational chemist ry and physics\, statistics\, and machine learning. The performance of MCM C methods has therefore become an important topic at the intersection of p robability theory and (computational) statistics: e.g.\, when the underlyi ng distribution one is trying to sample from becomes sufficiently complex\ , convergence speed and/or the cost per iteration becomes an issue for mos t MCMC methods. \n\nThe analysis\, and subsequently design\, of MCMC metho ds has to a large degree relied on classical tools used to determine the s peed of convergence of Markov chains\, e.g.\, mixing times\, spectral gap and functional inequalities (Poincaré\, log-Sobolev). An alternative aven ue is to use the theory of large deviations for empirical measures. In thi s talk I will first give a general outline of this approach to analysing M CMC methods\, along with some recent examples. I will then consider the sp ecific case of the Metropolis-Hastings algorithm\, the most classical amon gst all MCMC methods and a foundational building block for many more advan ced methods. Despite the simplicity of this method\, it turns out that the theoretical analysis of it is still a rich area\, and from the large devi ation perspective it is surprisingly difficult to treat. As a first step w e show a large deviation principle for the underlying Markov chain\, exten ding the celebrated Donsker-Varadhan theory. Time permitted I will also di scuss ongoing and future work on using this result for better understandin g of both the Metropolis-Hastings method and more advanced methods\, such as approximate Bayesian computation (ABC-MCMC) and the Metropolis-adjusted Langevin algorithm (MALA).\n\nThe talk will be self-contained and no prio r knowledge of either MCMC methods or large deviations is required.\n\nThe talk is primarily based on join work with Federica Milinanni (KTH).\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/42/ END:VEVENT BEGIN:VEVENT SUMMARY:Vincent Szolnoky (Chalmers University of Technology & University o f Gothenburg) DTSTART:20240306T121500Z DTEND:20240306T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/45 DESCRIPTION:Title: On the Interpretability of Regularisation for Neural Networks Thr ough Model Gradient Similarity\nby Vincent Szolnoky (Chalmers Universi ty of Technology & University of Gothenburg) as part of Gothenburg statist ics seminar\n\nLecture held in MVL14.\n\nAbstract\nMost complex machine le arning and modelling techniques are prone to over-fitting and may subseque ntly generalise poorly to future data. Artificial neural networks are no d ifferent in this regard and\, despite having a level of implicit regularis ation when trained with gradient descent\, often require the aid of explic it regularisers. We introduce a new framework\, Model Gradient Similarity (MGS)\, that (1) serves as a metric of regularisation\, which can be used to monitor neural network training\, (2) adds insight into how explicit re gularisers\, while derived from widely different principles\, operate via the same mechanism underneath by increasing MGS\, and (3) provides the bas is for a new regularisation scheme which exhibits excellent performance\, especially in challenging settings such as high levels of label noise or l imited sample sizes.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/45/ END:VEVENT BEGIN:VEVENT SUMMARY:Selma Tabakovic (Chalmers University of Technology & University of Gothenburg) DTSTART:20240313T121500Z DTEND:20240313T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/46 DESCRIPTION:Title: AI-driven sepsis care: early detection and personalized treatment \nby Selma Tabakovic (Chalmers University of Technology & University o f Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nSepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection\, and remains a leading cause of death in intensive care units worldwide. An optimal treatment strategy is still unknown\, leading to a significant variability in sepsis treatme nt with poorer outcomes.\n\nRecently\, deep reinforcement learning has sho wn promise as a decision-aiding tool for the administration of intravenous fluids and vasopressors to septic patients. However\, these models are li mited in their ability to accommodate the entire range from high-risk to l ow-risk patients\, and thus fail to provide personalized treatment recomme ndations.\n\nTo address this limitation\, in particular in the presence of heterogeneous patient groups or heterogeneous treatment responses\, we pr opose a Multi-Head Dueling Double Deep Q-Network (MH-DQN) model that incor porates patient characteristics to enable more personalized treatment reco mmendations. The MH-DQN model has multiple output layers\, each of which i s optimized for a specific patient profile. The model is trained using the Medical Information Mart for Intensive Care (MIMIC-III) database.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/46/ END:VEVENT BEGIN:VEVENT SUMMARY:Akash Sharma (Chalmers University of Technology & University of Go thenburg) DTSTART:20240320T121500Z DTEND:20240320T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/47 DESCRIPTION:Title: Sampling on manifolds via Langevin diffusion\nby Akash Sharma (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nWe d erive error bounds for sampling and estimation using a discretization of a n intrinsically defined Langevin diffusion on a compact Riemannian manifol d. Two estimators of linear functionals of invariant measure based on the discretized Markov process are considered: a time-averaging estimator and an ensemble-averaging estimator. Imposing no restrictions beyond a nominal level of smoothness on potential function\, first-order error bounds\, in discretization step size\, on the bias and variances of both estimators a re derived. We will also discuss conditions for extending analysis to the case of non-compact manifolds and different variants of the algorithm. We will present numerical illustrations with distributions on the manifolds o f positive and negative curvature which verify the derived bounds.\n\nJoin t work with Karthik Bharath (University of Nottingham)\, Alexander Lewis ( University of Gottingen) and Michael Tretyakov (University of Nottingham)\ n LOCATION:https://stable.researchseminars.org/talk/gbgstats/47/ END:VEVENT BEGIN:VEVENT SUMMARY:Nathan Gillot (University of Lorraine) DTSTART:20240403T111500Z DTEND:20240403T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/48 DESCRIPTION:Title: Modelling of the spread of a pathogen in coniferous forests and m odelling for cosmological data characterization\nby Nathan Gillot (Uni versity of Lorraine) as part of Gothenburg statistics seminar\n\nLecture h eld in MVL14.\n\nAbstract\nAs stated in the title\, this presentation will be divided into two parts. The first will deal with work on epidemiologic al data in a coniferous forest. We carried out modeling by considering pro bability laws on a lattice\, used the Gibbs sampler for simulation and use d two parameter approximation methods for these models: pseudo-likelihood and a Bayesian method\, the ABC Shadow algorithm. The second part of the t alk will focus on cosmological data. This time\, modeling will be done by spatial point processes\, simulation by the Metropolis Hastings algorithm and inference again by the ABC Shadow algorithm. The aim of this presentat ion is to give an idea of the tools used for the modelling\, simulation an d inference process for these two projects.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/48/ END:VEVENT BEGIN:VEVENT SUMMARY:Sergei Zuyev (Chalmers) DTSTART:20240417T111500Z DTEND:20240417T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/49 DESCRIPTION:Title: Training Bayesian neural networks with measure optimisation algor ithms\nby Sergei Zuyev (Chalmers) as part of Gothenburg statistics sem inar\n\nLecture held in MVL14.\n\nAbstract\nOn a high abstraction level\, a Bayesian neural network (BNN) can be seen as a function\nof input data a nd their prior probability distribution which yields\,\namong other output s\, their estimated posterior probability\ndistribution. This distribution is a result of optimisation of a\nchosen score function aiming to favour these probability distributions\nwhich describe best the observed data and take into account the prior\ndistribution.\n\nInstead of constraint optim isation over the simplex of probability distributions\, it is typical to\n map this simplex into Euclidean space\, for example with Softmax function or its variants\, and then do optimisation in the whole\nspace without co nstraints. It is\, however\, widely acknowledged that such mapping often s uffers from undesirable properties for optimisation and\nstability of the algorithms. To counterfeit this\, a few regularisation procedures have bee n proposed in the literature.\n\nInstead of trying to modify the mapping approach\, we suggest\nturning back to optimisation on the original simple x using recently\ndeveloped algorithms for constrained optimisation of fun ctionals of measures. \nWe demonstrate that our algorithms run tens times faster\nthan the standard algorithms involving softmax mapping and lead to exact solutions rather than to their approximations.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/49/ END:VEVENT BEGIN:VEVENT SUMMARY:Radu Stoica (Université de Lorraine) DTSTART:20240424T111500Z DTEND:20240424T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/51 DESCRIPTION:Title: Approximated inference for marked Gibbs point process\nby Rad u Stoica (Université de Lorraine) as part of Gothenburg statistics semina r\n\nLecture held in MVL14.\n\nAbstract\nParameter estimation for point pr ocesses is achieved via solving optimisation problems built using general strategies. Three well established strategies are enumerated. The first co nsists of considering contrast fuctions based on summary statistics. The s econd one uses the pseudo-likelihood. And the third approximates the likel ihood function via Monte Carlo procedures. Each of these techniques has kn own advantages and drawbacks (Moler and Waagepetersen 2004\, van Lieshout 2001\, 2019).\n\nSampling point process posterior densities is an inferenc e approach deeply intertwinned wih the previous one\, since it allows simu ltaneous parameter estimation and statistical tests based on observations. The auxiliary variable method (Moller et al.\,2006) gives the mathematica l solution to this problem\, while pointing out the difficulties of its pr actical implementation due to poor mixing. The exchange algorithm proposed by (Murray et al. 2006)\, (Caimo and Friel\, 2011) proposes a solution fo r the poor mixing induced by the auxiliary variable method. As its predece ssor it requires exact simulation for the sampling of the auxiliary variab le. This is not really a drawback\, but it may explode the computational t ime for models exhibiting strong interactions (van Lieshout and Stoica\, 2 006). \n\nThis talk presents the approximate ABC Shadow and SSA methods as complementary inference methods to the ones based on posterior density sa mpling. These methods do not require exact simulation\, while providing th e necessary theoretical control. The derived algorithms are applied on dat a from several application domains such as astronomy\, geosciences and ne twork sciences (Stoica et al.\,17)\, (Stoica et al.\,21)\, (Hurtado et al. \,21)\, (Laporte et al.\,22).\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/51/ END:VEVENT BEGIN:VEVENT SUMMARY:Robert Berman (Chalmers University of Technology & University of G othenburg) DTSTART:20240508T111500Z DTEND:20240508T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/53 DESCRIPTION:Title: Emergent complex geometry\nby Robert Berman (Chalmers Univers ity of Technology & University of Gothenburg) as part of Gothenburg statis tics seminar\n\nLecture held in MVL14.\n\nAbstract\nA recurrent theme in g eometry is the quest for canonical metrics on a given manifold X. The prot otypical case is when X is a compact orientable two-dimensional surface. S uch a manifold can be endowed with a metric of constant curvature\, which is uniquely determined by a fixing a complex structure on X. However\, fro m a physical point of view\, geometrical shapes - as we know them from eve ryday experience - are\, of course\, not fundamental physical entities. Th ey merely arise as macroscopic emergent features of ensembles of microscop ic point particles in the limit as the number N of particles tends to infi nity. This leads one to wonder if there is a canonical random point proces s on a given complex manifold X\, from which a canonical metrics emerges a s the number N of points tends to infinity? This is\, indeed\, the case\, when X is a complex algebraic hypersurface of any dimension\, as explained in the present talk. In this case the emerging metrics in question have c onstant Ricci curvature. More precisely\, they are Kähler-Einstein metric s. The talk is aimed to be non-technical and no previous background in com plex geometry is required.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/53/ END:VEVENT BEGIN:VEVENT SUMMARY:Umberto Picchini (Chalmers University of Technology & University o f Gothenburg) DTSTART:20240515T111500Z DTEND:20240515T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/54 DESCRIPTION:Title: Fast\, lightweight and semi-amortised simulation-based inference< /a>\nby Umberto Picchini (Chalmers University of Technology & University o f Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nBayesian inference for complex models with an intracta ble likelihood can be tackled using algorithms performing many calls to co mputer simulators. These approaches are collectively known as "simulation- based inference" (SBI). Recent SBI methods use neural-conditional-estimati on\, that is neural networks are employed to provide approximations to the likelihood function or the posterior distribution of model parameters. Wh ile neural-based posterior and likelihood estimation methods have produced exceptionally flexible inference strategies\, these can be computationall y intensive to run and have a non-negligible impact on energy expenditure and memory requirements. In this work\, rather than using neural networks we propose more "frugal" strategies that display state-of-art inference qu ality\, while being able to run with limited resources\, being much faster to train and exhibiting a much smaller computational footprint. We invest igate structured mixtures of probability distributions and design a new SB I method named Sequential Mixture Posterior and Likelihood Estimation (SeM PLE). SeMPLE learns closed-form approximations for both the posterior $p( θ|y)$ and the likelihood $p(y|θ)$ from the same training data\, using Ga ussian mixture models that can be efficiently learned.\nWe show favorable results for a variety of stochastic models (including SDEs and Markov jump processes)\, also in presence of multimodal posteriors. \n\nThe talk will be approachable for the uninitiated audience\, while novel results will b e of interest for the experienced audience.\n\nJoint work with Henrik Häg gström\, Pedro L. C. Rodrigues\, Geoffroy Oudoumanessah and Florence Forb es\, https://arxiv.org/abs/2403.07454\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/54/ END:VEVENT BEGIN:VEVENT SUMMARY:Adrien Corenflos (University of Warwick) DTSTART:20240529T111500Z DTEND:20240529T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/57 DESCRIPTION:Title: Particle-MALA and Particle-mGrad: Gradient-based MCMC methods for high-dimensional state-space models\nby Adrien Corenflos (University of Warwick) as part of Gothenburg statistics seminar\n\nLecture held in MV L14.\n\nAbstract\nState-of-the-art methods for Bayesian inference in state -space models are (a) conditional sequential Monte Carlo (CSMC) algorithms \; (b) sophisticated 'classical' MCMC algorithms like MALA\, or mGRAD from Titsias and Papaspiliopoulos (2018). The former propose N particles at ea ch time step to exploit the model's 'decorrelation-over-time' property and thus scale favourably with the time horizon\, T\, but break down if the d imension of the latent states\, D\, is large. The latter leverage gradient /prior-informed local proposals to scale favourably with D but exhibit sub -optimal scalability with T due to a lack of model-structure exploitation. We introduce methods which combine the strengths of both approaches. The first\, Particle-MALA\, spreads N particles locally around the current sta te using gradient information\, thus extending MALA to T>1 time steps and N>1 proposals. The second\, Particle-mGRAD\, additionally incorporates (co nditionally) Gaussian prior dynamics into the proposal\, thus extending th e mGRAD algorithm. We prove that Particle-mGRAD interpolates between CSMC and Particle-MALA\, resolving the 'tuning problem' of choosing between CSM C (superior for highly informative prior dynamics) and Particle-MALA (supe rior for weakly informative prior dynamics). We similarly extend other 'cl assical' MCMC approaches like auxiliary MALA\, aGRAD\, and preconditioned Crank-Nicolson-Langevin (PCNL). In experiments\, our methods substantially improve upon both CSMC and sophisticated `classical' MCMC approaches for both highly and weakly informative prior dynamics.\n\nTL\;DR: We aim to so lve the curse of dimensionality in state-space model inferences by combini ng the nice property (in time) of conditional particle filtering methods\, with the nice property (in space) of MALA and other gradient-based algori thms.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/57/ END:VEVENT BEGIN:VEVENT SUMMARY:Helga Olafsdottir (Chalmers University of Technology & University of Gothenburg) DTSTART:20240821T111500Z DTEND:20240821T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/62 DESCRIPTION:Title: Scoring rule inference for spatial statistics based on cross-vali dation\nby Helga Olafsdottir (Chalmers University of Technology & Univ ersity of Gothenburg) as part of Gothenburg statistics seminar\n\nLecture held in MVF21 (sic!).\n\nAbstract\nAlthough scoring rules are traditionall y aimed at model evaluation\, they have also successfully been used for mo del inference. We propose parameter inference of spatial models through a leave-one-out cross-validation approach (LOOS)\, where the predictive abil ity is optimised instead of the likelihood. The approach is studied for di fferent Gaussian spatial models. For Gaussian models with sparse precision matrices\, such as spatial Markov models\, the approach results in fast c omputations compared to the likelihood approach. Moreover\, the approach a llows affecting the robustness to outliers and sensitivity to non-stationa rity. Applying the LOOS to ERA5 temperature reanalysis data for the contig uous United States and the average July temperature for the years 1940 to 2023 resulted in estimates with better predictive performance than the max imum likelihood in a fraction of the computation time.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/62/ END:VEVENT BEGIN:VEVENT SUMMARY:Erik Jansson (Chalmers University of Technology and University of Gothenburg) DTSTART:20240828T111500Z DTEND:20240828T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/63 DESCRIPTION:Title: Sampling non-stationary Gaussian random fields on hypersurfaces u sing surface finite element methods\nby Erik Jansson (Chalmers Univers ity of Technology and University of Gothenburg) as part of Gothenburg stat istics seminar\n\nLecture held in MVL14.\n\nAbstract\nIn the middle of the previous century\, Peter Whittle demonstrated that Whittle–Matérn Gaus sian\nrandom fields on Euclidean domains can be obtained as solutions to f ractional elliptic stochastic\npartial differential equations (SPDEs). The SPDE–random field connection can be leveraged to\ngenerate random field s on other domains\, such as curves or surfaces\, by solving an SPDE on\nt hat domain. Selecting a differential operator with variable coefficients\, we obtain a flexible\nclass of non-stationary random fields. We consider how the computational technique of surface\nfinite elements can be utilize d to sample random fields on surfaces and briefly discuss how\nstrong erro r bounds are obtained using complex analysis and operator theory. \nThis talk is based on joint work with Annika Lang and Mike Pereira.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/63/ END:VEVENT BEGIN:VEVENT SUMMARY:Peter Guttorp (Norsk Regnesentral) DTSTART:20241113T121500Z DTEND:20241113T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/64 DESCRIPTION:Title: Water is warmer than air\, so why do we use sea surface temperatu re to estimate global temperature?\nby Peter Guttorp (Norsk Regnesentr al) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\n Abstract\nIn the study of global climate\, ocean temperature estimates use sea surface temperature (SST) anomalies instead of marine atmospheric tem perature (MAT) anomalies. A key question is to ask what biases result from this choice. In this talk we employ hierarchical statistical models to in vestigate spatial-temporal differences between SST and MAT anomalies in th e tropical Pacific. The analysis uses observations from the Tropical Atmos phere Ocean (TAO) buoy network. We use a spatio-temporal modeling approach that accounts for missing data in the observation network\, and allows fo r full uncertainty quantification. We also compare our results to another analysis that uses data from the European Center for Medium Range Weather Forecasting fifth generation reanalysis product (ERA5). We find no indica tion of bias or trend in replaciing MAT by SST in calculating global tempe rature anomalies.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/64/ END:VEVENT BEGIN:VEVENT SUMMARY:Satish Iyengar (Department of Statistics\, University of Pittsburg h\, Pittsburgh\, PA\, USA) DTSTART:20240925T111500Z DTEND:20240925T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/65 DESCRIPTION:Title: A clustering problem arising in psychiatry\nby Satish Iyengar (Department of Statistics\, University of Pittsburgh\, Pittsburgh\, PA\, USA) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\ nAbstract\nCurrent psychiatric diagnoses are based primarily on self-repor ted experiences. Unfortunately\, treatments for the diagnoses are not effe ctive for all patients. One hypothesized reason is that ``artificial group ing of heterogeneous syndromes with different pathophysiological mechanism s into one disorder.'' To address this problem\, the US National Institute of Mental Health instituted the Research Domain Criteria framework in 200 9. This research framework calls for integrating data from many levels of information: genes\, cells\, molecules\, circuits\, physiology\, behavior\ , and self-report. Clustering comes to the forefront as a key tool in this effort. In this talk\, I present a case study of the use of mixture model s to cluster older adults based on measures of sleep from three domains: d iary\, actigraphy\, and polysomnography. Challenges in this effort include the use of mixtures of skewed distributions\, a large number of potential clustering variables\, and seeking clinically meaningful solutions. We pr esent novel variable selection algorithms\, study them by simulation\, and demonstrate our methods on the sleep data. This work is joint with Meredi th Wallace.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/65/ END:VEVENT BEGIN:VEVENT SUMMARY:Linn Engström (KTH Royal Institute of Technology) DTSTART:20241023T111500Z DTEND:20241023T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/68 DESCRIPTION:Title: Computation of Robust Option Prices via Martingale Optimal Transp ort\nby Linn Engström (KTH Royal Institute of Technology) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nDurin g the last decade there has been a rapid development of methods for comput ationally addressing optimal transport problems\; motivated by application s within robust finance\, effort has also been made to generalize some of these techniques to problems equipped with an additional martingale constr aint. Computationally solving multi-marginal martingale optimal transport problems remains a challenging task though\, particularly for problems for mulated with a large number of marginals.\n\nIn this talk I will give a br ief introduction to the martingale optimal transport problem and motivate why it is interesting from a mathematical finance point of view\, before p resenting an efficient framework for solving a class of such multi-margina l problems computationally. The method combines the celebrated entropic re gularization with the exploitation of certain structures inherent in the p roblem\, enabling fast computation of the optimal dual variables. I will a lso provide some examples that demonstrates the utility of our method in t erms of computing model-independent bounds on the fair price of some exoti c options\, such as lookback options and Asian options. The talk is based on joint work with Sigrid Källblad and Johan Karlsson.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/68/ END:VEVENT BEGIN:VEVENT SUMMARY:Marcel Wienöbst (University of Lübeck) DTSTART:20241106T090000Z DTEND:20241106T094500Z DTSTAMP:20260404T095849Z UID:gbgstats/70 DESCRIPTION:Title: Linear-Time Algorithms for Front-Door Adjustment in Causal Graphs \nby Marcel Wienöbst (University of Lübeck) as part of Gothenburg st atistics seminar\n\nLecture held in MVL14.\n\nAbstract\nCausal effect esti mation from observational data is a fundamental task in\nempirical science s. It becomes particularly challenging when unobserved\nconfounders are in volved in a system. Front-door adjustment constitutes a\nclassic method th at allows identifying the causal effect even in the presence of\nlatent co nfounding by using observed mediators. This talk presents a recent\nalgori thmic result in this area\, namely a linear-time algorithm for finding a\n front-door adjustment set in a given causal graph. Its run-time is\nasympt otically optimal and improves on the previous state-of-the-art for this\nt ask by a factor that grows cubically in the number of variables. Beyond th is\nresult\, the presentation explores fundamental algorithmic tools and t echniques\nuseful for broader applications in causal inference.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/70/ END:VEVENT BEGIN:VEVENT SUMMARY:Sebastian Persson (University of Gothenburg) DTSTART:20241127T121500Z DTEND:20241127T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/72 DESCRIPTION:Title: PEtab.jl - Efficient parameter estimation for dynamic models\ nby Sebastian Persson (University of Gothenburg) as part of Gothenburg sta tistics seminar\n\nLecture held in MVL14.\n\nAbstract\nOrdinary differenti al equations (ODEs) are commonly used to model dynamic processes such as b iological networks. ODE models often contain unknown parameters that must be estimated from data. From a statistical viewpoint\, this estimation is typically performed by computing a maximum likelihood estimate\, which boi ls down to solving a nonlinear optimization problem. In simple cases\, the likelihood function can be easily coded using existing libraries in progr amming languages like Python and Julia. However\, for more complex scenari os—such as when the model includes events\, data is collected under vari ous simulation conditions\, or the model should be at a steady state at ti me zero—correctly coding a likelihood function becomes time-consuming an d error-prone. Moreover\, numerically fitting an ODE model to data can be computationally intensive\, potentially taking hours to days\, and the cho ice of ODE solver and gradient computation methods can drastically affect runtime. \n\n \nIn this talk\, I will discuss our software package PEtab.j l\, a Julia package for setting up parameter estimation problems for dynam ic models. I will cover how PEtab.jl simplifies parameter estimation workf lows and present extensive benchmark results on how the choice of gradient methods and ODE solvers affects runtime. Lastly\, I will discuss how mech anistic models can be complemented with data-driven neural-network models to address the shortcomings of each model type.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/72/ END:VEVENT BEGIN:VEVENT SUMMARY:Lembris Njotto (University of Dar es Salaam) DTSTART:20241120T121500Z DTEND:20241120T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/73 DESCRIPTION:Title: Spatio-temporal modeling of malaria cases in Tanzania\nby Lem bris Njotto (University of Dar es Salaam) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nMalaria continues to pose a significant global health challenge\, affecting approximately 200 millio n individuals annually and causing an estimated 600\,000 deaths worldwide. Environmental factors are key drivers of malaria transmission dynamics\, influencing disease patterns at local and regional scales. This talk focus es on data from Tanzania to explore the impact of climatic factors and vec tor control interventions on malaria incidence.\n\nUsing Standardized Inci dence Ratio (SIR) metrics and Bayesian spatio-temporal modeling approaches \, we analyze regionally aggregated monthly malaria cases\, stratified int o two age groups: children under five and individuals aged five years and above. The models incorporate a Conditional Autoregressive (CAR) structure to capture spatial dependencies\, a second-order random walk (RW2) for te mporal trends\, and independent and identically distributed (iid) random e ffects to account for unstructured spatial and temporal variability. Speci fic results on the influence of environmental factors\, including precipit ation and temperature\, on malaria cases will be presented during the talk \, highlighting their role in transmission dynamics and informing targeted intervention strategies.\n\nResults are not yet published\, please make t hem confidential.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/73/ END:VEVENT BEGIN:VEVENT SUMMARY:Martin Voigt Vejling (Aalborg University) DTSTART:20241121T100000Z DTEND:20241121T110000Z DTSTAMP:20260404T095849Z UID:gbgstats/74 DESCRIPTION:Title: Conformal multiple Monte Carlo testing with a view to spatial sta tistics\nby Martin Voigt Vejling (Aalborg University) as part of Gothe nburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nMonte Carl o tests are popular for their convenience\, as they allow the computation of valid p-values even when test statistics with known and tractable distr ibutions are unavailable. When performing multiple Monte Carlo tests\, it is essential to adjust the testing procedure to maintain control of the ty pe I error\, and some of such techniques pose requirements on the joint di stribution of the p-values\, for instance independence. A straightforward approach to get independent p-values\, is to compute the p-values for each hypothesis in parallel\, however\, this imposes a substantial computation al burden. We highlight in this work that the problem of testing multiple data samples against the same null hypothesis is an instance of conformal outlier detection. Leveraging this insight enables a more efficient multip le Monte Carlo testing procedure\, avoiding excessive simulations while st ill ensuring exact control over the false discovery rate. Through numerica l experiments on point patterns we investigate the performance of this pro posed conformal multiple Monte Carlo testing (CMMCTest) procedure.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/74/ END:VEVENT BEGIN:VEVENT SUMMARY:Aila Särkkä (Chalmers University of Technology & University of G othenburg) DTSTART:20241211T121500Z DTEND:20241211T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/75 DESCRIPTION:Title: Analysis of point patterns observed with errors: some examples\nby Aila Särkkä (Chalmers University of Technology & University of Got henburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14 .\n\nAbstract\nMany natural systems are observed as point patterns in time \, space\, or space and time. Examples include plant and cellular systems\ , animal colonies\, wildfires\, and galaxies. In practice\, the locations of the points are not always observed correctly. However\, in the point pr ocess literature\, little attention has been paid to the issue of errors i n the locations of points. In this talk\, we discuss how the observed poin t pattern may deviate from the actual point pattern\, review methods and m odels that exist to handle such deviations\, and give some examples of dat a observed with errors. \n\nBased on joint work with Peter Guttorp (Norweg ian Computing Center)\, Janine Illian (University of Glasgow)\, Joel Koste nsalo (Natural Resources Institute Finland (Luke)\, Mikko Kuronen (Luke)\, Mari Myllymäki (Luke)\, and Thordis Thorarinsdottir (University of Oslo) .\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/75/ END:VEVENT BEGIN:VEVENT SUMMARY:Henrik Häggström (Chalmers University of Technology & University of Gothenburg) DTSTART:20250122T121500Z DTEND:20250122T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/76 DESCRIPTION:Title: Simulation-based inference for stochastic nonlinear mixed-effects models with applications in systems biology\nby Henrik Häggström (C halmers University of Technology & University of Gothenburg) as part of Go thenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nWe prop ose a novel methodology for Bayesian inference in hierarchical mixed-effec ts models. We construct a framework that is highly scalable\, where amorti zed approximations to the likelihood and the parameters posterior are firs t obtained\, and these are rapidly refined for each individual dataset\, t o ultimately approximate the parameters posterior across many individuals. The framework introduced is easily trainable\, as it uses mixture of expe rts but without neural networks\, leading to parsimonious yet expressive s urrogate models of the likelihood and the posterior. The methodology is ex emplified via stochastic differential equation mixed-effects models\, that are highly relevant in systems biology\, but the methodology is general a nd can accommodate other types of stochastic and deterministic models. We compare our approximate inference with exact pseudomarginal inference and show that our methodology is fast and competitive.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/76/ END:VEVENT BEGIN:VEVENT SUMMARY:Julia Jansson (Chalmers University of Technology & University of G othenburg) DTSTART:20250109T083000Z DTEND:20250109T103000Z DTSTAMP:20260404T095849Z UID:gbgstats/77 DESCRIPTION:Title: Licentiate seminar: Statistical Properties of Point Process Learn ing for Gibbs Processes\nby Julia Jansson (Chalmers University of Tech nology & University of Gothenburg) as part of Gothenburg statistics semina r\n\nLecture held in Pascal\, Hörsalsvägen 1.\n\nAbstract\nThis thesis s tudies Point Process Learning (PPL)\, which is a novel statistical learnin g framework that uses point process cross-validation and point process pre diction errors\, and includes different hyperparameters. Specifically\, st atistical properties of PPL are explored\, in the context of Gibbs point p rocesses. Paper 1 demonstrates PPL’s advantages over pseudolikelihood\, which is a state-of-the-art parameter estimation method and a special case of Takacs- Fiksel estimation (TF)\, with particular focus on Gibbs hard-c ore processes. Paper 2 compares PPL to TF\, and shows that TF is a special case of PPL\, when the cross-validation scheme tends to leave-one-out cro ss-validation. In addition\, Paper 2 shows that for four common Gibbs mode ls\, namely Poisson\, hard-core\, Strauss and Geyer saturation processes\, one can choose hyperparameters so that PPL outperforms TF in terms of mea n square error.\n\nIn Paper 1 and 2\, parameter estimation with PPL is don e by minimizing loss functions\, while Paper 3 explores an alternative app roach to PPL\, namely estimating equations. Further\, statistical properti es of the parameter estimator are derived in Paper 3\, such as consistency and asymptotic normality for large samples\, as well as bias and variance for small samples. It is concluded that the estimating equation approach is not feasible for PPL\, whereby the original loss function-based approac h is preferred. Moving on\, Paper 3 then provides a theoretical foundation for the loss functions through an empirical risk formulation.\n\nTo concl ude\, PPL is shown to be a flexible and robust competitor to state-of-the- art methods for parameter estimation.\n\nRoom: Pascal\, Hörsalsvägen 1\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/77/ END:VEVENT BEGIN:VEVENT SUMMARY:Alexey Lindo (University of Glasgow) DTSTART:20250115T121500Z DTEND:20250115T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/78 DESCRIPTION:Title: Probability-Generating Function Kernels for Spherical Data\nb y Alexey Lindo (University of Glasgow) as part of Gothenburg statistics se minar\n\nLecture held in MVL14.\n\nAbstract\nIn this talk\, I will introdu ce the class of probability-generating function (PGF) kernels\, a novel ap proach to spherical data analysis. PGF kernels generalize radial basis fun ction (RBF) kernels and are supported on the unit hypersphere\, making the m well-suited for tasks involving spherical data. I will discuss their uni que properties\, demonstrate a semi-parametric learning algorithm for fitt ing these kernels\, and showcase their application in Gaussian processes a nd deep kernel learning. Through examples and comparisons\, I will highlig ht the advantages of PGF kernels over existing methods.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/78/ END:VEVENT BEGIN:VEVENT SUMMARY:Sara Hamis (Uppsala University) DTSTART:20250423T111500Z DTEND:20250423T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/79 DESCRIPTION:Title: Predicting and controlling cell systems that generate spatio-temp oral point patterns\nby Sara Hamis (Uppsala University) as part of Got henburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nRecent t echnological advances have resulted in a multitude of spatio-temporal cell imaging data. These can be translated into spatio-temporal point patterns in which points represent cells. Such data hold rich information about ho w cells act and interact\, much of which is not extractable through data a nalysis alone. Therefore\, to identify\, predict and control cell systems that generate spatio-temporal patterns\, we propose using two unified clas ses of mathematical models: spatio-temporal point processes (STPPs) and sp atial cumulant models (SCMs). SCMs are population models formulated by dif ferential equations that approximate the dynamics of two STPP-generated su mmary statistics: first-order spatial cumulants (densities)\, and second-o rder spatial cumulants (spatial covariances). In this talk\, I’ll demons trate that (1) SCMs can capture STPP-generated density dynamics\, even whe n mean-field population models (MFPMs) fail to do so\, and (2) SCM-informe d treatment strategies outperform MFPM-informed strategies in terms of inh ibiting population growths. Overall\, our work demonstrates that SCMs prov ide a promising framework in which to study ecological systems that genera te spatio-temporal point patterns in cell biology and beyond.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/79/ END:VEVENT BEGIN:VEVENT SUMMARY:Alvaro Köhn-Luque (University of Oslo) DTSTART:20250430T090000Z DTEND:20250430T094500Z DTSTAMP:20260404T095849Z UID:gbgstats/80 DESCRIPTION:Title: Phenotypic deconvolution of cancer cell populations\nby Alvar o Köhn-Luque (University of Oslo) as part of Gothenburg statistics semina r\n\nLecture held in MVL15.\n\nAbstract\nTumor heterogeneity is an importa nt driver of treatment failure in cancer\, as therapies often select for d rug-tolerant or drug-resistant cellular subpopulations that drive tumor gr owth and recurrence. Profiling the drug-response heterogeneity of tumor sa mples using traditional genomic deconvolution methods has yielded limited results\, due in part to the imperfect mapping between genomic variation a nd functional characteristics. In this seminar\, I will demonstrate how to leverage mechanistic population modeling to develop a statistical framewo rk for profiling phenotypic heterogeneity from standard drug-screen data o n bulk tumor samples. This approach allows us to reliably identify tumor s ubpopulations exhibiting differential drug responses and estimate their dr ug sensitivities and frequencies within the bulk population. I will discus s the advantages and disadvantages of using deterministic versus stochasti c birth-death population models. These methods are applied to syntheticall y generated cell populations\, mixed cell-line in vitro experiments\, and multiple myeloma patient samples.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/80/ END:VEVENT BEGIN:VEVENT SUMMARY:Calum Gabbutt (Imperial College London) DTSTART:20250507T111500Z DTEND:20250507T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/81 DESCRIPTION:Title: Timing copy number alterations in Barrett’s Oesophagus using hi erarchical Bayesian models\nby Calum Gabbutt (Imperial College London) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbs tract\nThe accumulation of somatic copy number alterations (CNAs) is a key genomic risk factor in the progression from Barrett’s Oesophagus (a pre -malignant condition\, BO) to oesophageal adenocarcinoma (OAC). However\, the timing and evolutionary dynamics of these CNAs have remained elusive. In this talk\, I will introduce CARBINE (Copy number AlteRation timing wit h Bayesian Inference and Neutral Evolution)\, a hierarchical Bayesian fram ework designed to infer the calendar-time occurrence of CNAs from deep who le-genome sequencing data. CARBINE leverages molecular clock signals and c lonal evolutionary theory to estimate patient-specific mutation rates\, cl onal growth dynamics\, and the timing of genomic events from single-timepo int samples. Using this method\, we find that critical alterations\, such as whole-genome doubling and TP53 inactivation\, often occur decades befor e cancer diagnosis\, including during early life\, and are followed by lon g periods of indolent clonal expansion. Furthermore\, we show that the rat e of CNA accumulation—estimated from single snapshots—outperforms over all burden as a predictor of progression to OAC. This new insight into the temporal evolution of BO underscores the potential of early-life genomic profiling to stratify cancer risk and informs strategies for early interve ntion.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/81/ END:VEVENT BEGIN:VEVENT SUMMARY:Rasmus Waagepetersen (Aalborg University) DTSTART:20250110T100000Z DTEND:20250110T104500Z DTSTAMP:20260404T095849Z UID:gbgstats/82 DESCRIPTION:Title: Point process approaches to study clustering of industry location s\nby Rasmus Waagepetersen (Aalborg University) as part of Gothenburg statistics seminar\n\nLecture held in MVL15.\n\nAbstract\nIn this talk we will discuss various point process approaches to study clustering of indus try locations. Industries (shops\, firms\, supermarkets\, factories...) ca n be of various types and clustering (or possibly repulsion) could happen within industries of the same type or between industries of different type s. We review a seminal contribution in spatial econometrics and discuss it s relation to recent semi-parametric point process models including semi-p arametric log Gaussian Cox processes and semi-parametric Markov point proc esses. For the semi-parametric models we in particular consider how parame ter estimates can be obtained using certain conditional composite likeliho ods.\n\nRoom: MVL15\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/82/ END:VEVENT BEGIN:VEVENT SUMMARY:Elizabeth Baker (DTU) DTSTART:20250611T111500Z DTEND:20250611T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/83 DESCRIPTION:Title: Conditioning diffusion processes with score matching methods\ nby Elizabeth Baker (DTU) as part of Gothenburg statistics seminar\n\nLect ure held in MVL14.\n\nAbstract\nIn stochastic optimal control and conditio nal generative modelling\, a central computational task is to modify a ref erence diffusion process to maximise a given terminal-time reward. Most ex isting methods require this reward to be differentiable\, using gradients to steer the diffusion towards favourable outcomes. However\, in many prac tical settings\, like diffusion bridges\, the reward is singular\, taking an infinite value if the target is hit and zero otherwise. We introduce a novel framework\, based on Malliavin calculus and path-space integration b y parts\, that enables the development of methods robust to such singulari ties. This allows our approach to handle a broad range of applications\, i ncluding classification\, diffusion bridges\, and conditioning without the need for artificial observational noise.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/83/ END:VEVENT BEGIN:VEVENT SUMMARY:Mattias Byléhn (Chalmers University of Technology & University of Gothenburg) DTSTART:20250528T111500Z DTEND:20250528T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/84 DESCRIPTION:Title: PhD defence: Hyperuniformity and hyperfluctuations for random mea sures on Euclidean and non-Euclidean spaces\nby Mattias Byléhn (Chalm ers University of Technology & University of Gothenburg) as part of Gothen burg statistics seminar\n\nLecture held in Euler.\n\nAbstract\nReserved sl ot because of https://www.chalmers.se/en/current/calendar/mv-doctoral-thes is-mattias-byhlen/\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/84/ END:VEVENT BEGIN:VEVENT SUMMARY:Vincent Molin (Chalmers University of Technology & University of G othenburg) DTSTART:20250416T111500Z DTEND:20250416T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/85 DESCRIPTION:Title: Controlled stochastic processes for simulated annealing\nby V incent Molin (Chalmers University of Technology & University of Gothenburg ) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAb stract\nSimulated annealing solves optimization problems by means of a ran dom walk in an energy landscape based on the objective function and a temp erature parameter. By slowly decreasing the temperature\, the algorithm co nverges to the global optimal solution\, also for nonconvex functions. How ever\, if the temperature is decreased too quickly\, this procedure often gets stuck in local minima. To overcome this\, we here present a new persp ective on simulated annealing. More precisely\, we consider the cooling la ndscape as a curve of probability measures and prove that there exists a m inimal norm velocity field which solves the continuity equation. The latte r is a differential equation which governs the evolution of the aforementi oned curve. The solution is the weak gradient of an integrable function\, which is in line with the interpretation of the velocity field as a deriva tive of optimal transport maps. We also show that controlling stochastic a nnealing processes by superimposing this velocity field would allow them t o follow arbitrarily fast cooling schedules. Based on these findings\, we design novel interacting particle based optimization methods\, convergent optimal transport based approximations to this control\, that accelerate s imulated annealing processes. This acceleration behavior is also validated on a number of numerical experiments.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/85/ END:VEVENT BEGIN:VEVENT SUMMARY:Adélie Erard (Université Paris Cité) DTSTART:20250319T121500Z DTEND:20250319T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/86 DESCRIPTION:Title: A method for estimating population growth at a local scale. An ap plication using French breeding bird surveys data.\nby Adélie Erard ( Université Paris Cité) as part of Gothenburg statistics seminar\n\nLectu re held in MVL14.\n\nAbstract\nEstimating population growth at a local sca le is crucial for understanding ecological dynamics and informing conserva tion efforts. In this work\, we propose a novel methodology that uses mark ed point processes to model bird populations as a spatial process of unkno wn intensity influenced by environmental factors. The population is repres ented as a point process $\\mathcal{P}$\, while observations are obtained through a thinning mechanism using a homogeneous Poisson birth-and-death p rocess $\\mathcal{O}_t$. This approach accounts for both spatial dependenc e\, such as Cox or Gibbs processes\, and temporal variations by focusing o n intersections of observed areas over consecutive time points.\n\n\nBy ap plying stabilization theory\, we demonstrate convergence properties and en sure robust local estimations of population variation. The ability to esti mate population dynamics at fine spatial scales distinguishes this approac h\, making it particularly suited to ecological applications.\n\nAn implem entation using French breeding bird survey data highlights its potential t o capture localized trends and advance biodiversity monitoring.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/86/ END:VEVENT BEGIN:VEVENT SUMMARY:Nils Grimbeck (Chalmers University of Technology & University of G othenburg) DTSTART:20250527T083000Z DTEND:20250527T091500Z DTSTAMP:20260404T095849Z UID:gbgstats/87 DESCRIPTION:Title: What do we see when we look at art? - Modelling of eye-movements in art perception\nby Nils Grimbeck (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistics seminar\n\nL ecture held in MVL14.\n\nAbstract\nEye movements during art perception hav e been extensively studied over the past century\, as they provide insight s into perceptual\, evaluative\, and cognitive processes. Although several theoretical frameworks have been proposed\, it is only recently that spat ial statistics have begun to explore gaze patterns\, specifically by model ling fixation locations as spatial point patterns arising from spatio-temp oral point processes. Inspired by the simple model of eye-movements propos ed by Ylitalo et al. in 2016\, we propose a stochastic model that incorpor ates both the semi-conscious transitions between regions of interest (ROIs ) in a painting\, as well as the semi-random eye movements that occur whil e registering visual information within these regions during the first 30 seconds of art perception. \n \n\nUsing eye-tracking data from twenty s ubjects on six paintings\, we apply mean-shift clustering to identify ROIs in each painting based on the intensity of fixation points. A Markov chai n is subsequently used to model the transitions between these regions and based on the model proposed by Ylitalo et al. we use the estimated intensi ty and saccade length distribution to model the placement of fixations wit hin each ROI. Using this modelling approach\, we analyse the dynamics of e ye movements during the initial 30 seconds of art perception and to assess the robustness of our modelling assumptions across six diverse paintings and artistic styles.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/87/ END:VEVENT BEGIN:VEVENT SUMMARY:Marco Tarantino (University of Palermo) DTSTART:20250326T121500Z DTEND:20250326T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/88 DESCRIPTION:Title: Using a Neural Network approach and Starspots dependent models to predict Effective Temperatures and Ages of young stars\nby Marco Tara ntino (University of Palermo) as part of Gothenburg statistics seminar\n\n Lecture held in MVL14.\n\nAbstract\nThis study presents a statistical appr oach to accurately predict the effective temperatures of pre-main sequence stars\, which are necessary for determining stellar ages using the isochr one methodology and cutting-age starspots-dependent models. By training a Neural Network model on high-quality spectroscopic temperatures from the G aia-ESO Survey as the response variable\, and using photometric data from Gaia DR3 and 2MASS catalogs as explanatory variables\, we implemented a me thodology to accurately derive the effective temperatures of much larger p opulations of stars for which only photometric data are available. The mod el demonstrated robust performance for low-mass stars with temperatures be low 7000 K\, including young stars\, the primary focus of this work. Pred icted temperatures were employed to construct Hertzsprung-Russell diagrams and to predict stellar ages of different young clusters and star forming regions through isochrone interpolation\, achieving excellent agreement wi th spectroscopic-based ages and literature values derived from model-indep endent methods like lithium equivalent widths. The inclusion of starspot e volutionary models improved the age predictions\, providing a more accurat e description of stellar properties. Additionally\, the results regarding the effective temperature and age predictions of the young clusters provid e evidences of the presence of intrinsic age spreads in the youngest clust ers\, suggesting multiple formation events over time.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/88/ END:VEVENT BEGIN:VEVENT SUMMARY:Jeffrey Steif (Chalmers University of Technology & University of G othenburg) DTSTART:20250903T111500Z DTEND:20250903T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/89 DESCRIPTION:Title: Is your coin rational?\nby Jeffrey Steif (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistic s seminar\n\nLecture held in MVL14.\n\nAbstract\nOne tosses a coin with an unknown parameter p a large number of times and then you have to guess wh ether p is rational or irrational. Can you do it? The answer is related to various things such as\nthe so-called Baire Category Theorem as well as w ell-approximability of irrationals by rationals \nin elementary number the ory.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/89/ END:VEVENT BEGIN:VEVENT SUMMARY:Ruben Seyer (Chalmers University of Technology & University of Got henburg) DTSTART:20250521T111500Z DTEND:20250521T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/90 DESCRIPTION:Title: Creating non-reversible rejection-free samplers by rebalancing sk ew-balanced Markov jump processes\nby Ruben Seyer (Chalmers University of Technology & University of Gothenburg) as part of Gothenburg statistic s seminar\n\nLecture held in MVL14.\n\nAbstract\nMarkov chain sampling met hods form the backbone of modern computational statistics. However\, many popular methods are prone to random walk behavior\, i.e.\, diffusion-like exploration of the sample space\, leading to slow mixing that requires int ricate tuning to alleviate. Non-reversible samplers can resolve some of th ese issues. We introduce a device that turns jump processes that satisfy a skew-detailed balance condition for a reference measure into a process th at samples a target measure that is absolutely continuous with respect to the reference measure. The resulting sampler is rejection-free\, non-rever sible\, and continuous-time. As an example\, we apply the device to Hamilt onian dynamics discretized by the leapfrog integrator\, resulting in a rej ection-free non-reversible continuous-time version of Hamiltonian Monte Ca rlo (HMC). We prove the geometric ergodicity of the resulting sampler unde r certain convexity conditions\, and demonstrate its qualitatively differe nt behavior to HMC through numerical examples.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/90/ END:VEVENT BEGIN:VEVENT SUMMARY:Philip Gerlee (Chalmers University of Technology & University of G othenburg) DTSTART:20250618T111500Z DTEND:20250618T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/91 DESCRIPTION:Title: Evaluation of respiratory disease hospitalisation forecasts using synthetic outbreak data\nby Philip Gerlee (Chalmers University of Tec hnology & University of Gothenburg) as part of Gothenburg statistics semin ar\n\nLecture held in MVL14.\n\nAbstract\nForecasts of hospitalisations of infectious diseases play an important role for allocating healthcare reso urces during epidemics and pandemics. Large-scale analysis of model foreca sts during the COVID-19 pandemic has shown that the model rank distributio n with respect to accuracy is heterogeneous and that ensemble forecasts ha ve the highest average accuracy. Building on that work we generated a maxi mally diverse synthetic dataset of 324 different hospitalisation time-seri es that correspond to different disease characteristics and public health responses. We evaluated forecasts from 14 component models and 6 different ensembles. Our results show that component model accuracy was heterogeneo us and varied depending on the current rate of disease transmission. Going from 7 day to 14 day forecasts mechanistic models improved in relative ac curacy compared to statistical models. A novel adaptive ensemble method ou tperforms all other ensembles\, but is closely followed by a median ensemb le. We also investigated the relationship between ensemble error and varia bility of component forecasts and show that the coefficient of variation i s predictive of future error. Lastly\, we validated the results on data fr om the COVID-19 pandemic in Sweden. Our findings have the potential to imp rove epidemic forecasting\, in particular the ability to assign confidence to ensemble forecasts at the time of prediction based on component foreca st variability.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/91/ END:VEVENT BEGIN:VEVENT SUMMARY:Fima Klebaner (Monash University) DTSTART:20250710T111500Z DTEND:20250710T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/92 DESCRIPTION:Title: Emergence of populations. H(W) theory.\nby Fima Klebaner (Mon ash University) as part of Gothenburg statistics seminar\n\nLecture held i n MVL15.\n\nAbstract\nWe study how populations emerge when starting with j ust a few individuals\, maybe only one\, and then growing to its (large) c arrying capacity K. We prove an old conjecture and suggest new approximati ons.\n\nThe talk is based on a number of papers with: Andrew Barbour\, Pav el Chigansky\, Peter Jagers\, Kais Hamza\, and PhD students Jeremy Baker a nd Naor Bauman.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/92/ END:VEVENT BEGIN:VEVENT SUMMARY:Oskar Allerbo (KTH Royal Institute of Technology) DTSTART:20251008T111500Z DTEND:20251008T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/93 DESCRIPTION:Title: Is supervised learning really that different from unsupervised?\nby Oskar Allerbo (KTH Royal Institute of Technology) as part of Gothen burg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nWe demonstr ate how supervised learning can be decomposed into a two-stage procedure\, where (1) all model parameters are selected in an unsupervised manner\, a nd (2) the outputs y are added to the model\, without changing the paramet er values. This is achieved by a new model selection criterion that - in c ontrast to cross-validation - can be used also without access to y. For li near ridge regression\, we bound the asymptotic out-of-sample risk of our method in terms of the optimal asymptotic risk. We also demonstrate on rea l and synthetic data that versions of linear and kernel ridge regression\, smoothing splines\, and neural networks\, which are trained without acces s to y\, perform similarly to their standard y-based counterparts. Hence\, our results suggest that the difference between supervised and unsupervis ed learning is less fundamental than it may appear.\nJoint work with Thoma s B. Schön.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/93/ END:VEVENT BEGIN:VEVENT SUMMARY:Filip Tronarp (Lund University) DTSTART:20251119T121500Z DTEND:20251119T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/94 DESCRIPTION:Title: A Recursive Theory of Variational State Estimation: The Dynamic P rogramming Approach\nby Filip Tronarp (Lund University) as part of Got henburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nIn this talk\, we discuss the variational inference problem in partially observed Markov processes from the dynamic programming perspective. \nThis leads to a backward and a forward recursion for certain value functionals\, which are closely connected to the corresponding recursions from classical Bayes ian state estimation theory. Namely\, the backward value functional is a l ower bound on the "backward filter" and the forward value functional is a lower bound on the unnormalized filtering density. The two recursions can also be combined yielding a variational two-filter formula.\nWhat results is a variational state estimation theory that is completely analogous to t he classical Bayesian state estimation theory. \nThe theory is applied to a jump Gauss-Markov regression problem\, where closed form solutions to th e value functional recursions can be obtained.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/94/ END:VEVENT BEGIN:VEVENT SUMMARY:Johannes Borgqvist (Chalmers) DTSTART:20251015T111500Z DTEND:20251015T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/95 DESCRIPTION:Title: HeMiTo-dynamics: a characterization of mammalian prion toxicity u sing non-dimensionalization\, linear stability and perturbation analyses\nby Johannes Borgqvist (Chalmers) as part of Gothenburg statistics semi nar\n\nLecture held in MVL14.\n\nAbstract\nPrion-like proteins play crucia l parts in biological processes in organisms ranging from yeast to humans. For instance\, many neurodegenerative diseases are believed to be caused by the production of prion-like proteins in neural tissue. As such\, under standing the dynamics of prion-like protein production is a vital step tow ard treating neurodegenerative disease. Mathematical models of prion-like protein dynamics show great promise as a tool for predicting disease traje ctories and devising better treatment strategies for prion-related disease s. Herein\, we investigate a generic model for prion-like dynamics consist ing of a class of non-linear ordinary differential equations (ODEs)\, esta blishing constraints through a linear stability analysis that enforce the expected properties of mammalian prion-like toxicity. Furthermore\, we ide ntify that prion toxicity evolves through three distinct phases for which we provide analytical descriptions using perturbation analyses. Specifical ly\, prion-toxicity is initially characterized by the healthy phase\, wher e the dynamics are dominated by the healthy form of prions\, thereafter th e system enters the mixed phase\, where both healthy and toxic prions inte ract\, and lastly\, the system enters the toxic phase\, where toxic prions dominate\, and we refer to these phases as HeMiTo-dynamics. These finding s hold the potential to aid researchers in developing precise mathematical models for prion-like dynamics\, enabling them to better understand under lying mechanisms and devise effective treatments for prion-related disease s.\n\nAt this point in time\, the work has been solely focused on analysin g a class of mathematical models of prion diseases. Next\, the plan is to start two new projects involving experimental data from medical collaborat ors. In short\, these projects involve identifying an unknown conversion f unction in our class of prion models using time series data in combination with physics informed neural networks\, as well as spatial modelling of h ow prions are distributed over time in diseased brains. The main aim of th is talk is to start a discussion about these collaboration projects\, and any input would be greatly appreciated.\n\nThe slide-based presentation is made in beamer\, and I will bring my own laptop to the presentation.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/95/ END:VEVENT BEGIN:VEVENT SUMMARY:Kasper Bågmark (Chalmers) DTSTART:20260211T121500Z DTEND:20260211T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/96 DESCRIPTION:Title: High-dimensional Bayesian filtering through deep density approxim ation\nby Kasper Bågmark (Chalmers) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nIn this work\, we benchmark two recently developed deep density methods for nonlinear filtering. Star ting from the Fokker--Planck equation with Bayes updates\, we model the fi ltering density of a discretely observed SDE. The two filters: the deep sp litting filter and the deep BSDE filter\, are both based on Feynman--Kac f ormulas\, Euler--Maruyama discretizations and neural networks. The two met hods are extended to logarithmic formulations providing sound and robust i mplementations in increasing state dimension. Comparing to the classical p article filters and ensemble Kalman filters\, we benchmark the methods on numerous examples. In the low-dimensional examples the particle filters wo rk well\, but when we scale up to a partially observed $100$-dimensional L orenz-96 model the particle-based methods fail and the logarithmic deep de nsity method prevails. In terms of computational efficiency\, the deep den sity methods reduce inference time by roughly two to five orders of magnit ude relative to the particle-based filters.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/96/ END:VEVENT BEGIN:VEVENT SUMMARY:Sandra Barman (RISE Research Institutes of Sweden) DTSTART:20260204T121500Z DTEND:20260204T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/97 DESCRIPTION:Title: Correlating structure and material properties in soft materials\nby Sandra Barman (RISE Research Institutes of Sweden) as part of Gothe nburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nThis talk focuses on how statistics and machine learning can be used to correlate th e nano- and microstructure of soft materials with their material propertie s. We work with different application areas where this is relevant\, inclu ding packaging and barrier materials\, hygiene products\, pharmaceuticals\ , and food.\n\nTo develop models that map the relationship between a mater ial’s structure and its functional properties\, we combine:\n\n
    \n < li>methods for material imaging to understand what the structure looks lik e\, ranging from indirect methods such as X-ray scattering to direct imagi ng in 2D and 3D\, with or without a time component\,\n
  1. models fo r replicating and exploring 3D material structure using spatial statistics and generative AI\,
    2. \n
  2. numerical simulation of functional proper ties such as fluid and gas transport\, and
    3. \n
  3. statistical and mac hine learning models that connect structure to functional properties.
    4. \n
\n\nI will in this talk give an overview of some ongoing projects w hich are done in collaboration between RISE\, the Department of Mathematic al Sciences and Department of Physics at Chalmers\, Chalmers Industritekni k\, and industrial partners such as Tetra Pak\, AstraZeneca\, and Essity.\ n LOCATION:https://stable.researchseminars.org/talk/gbgstats/97/ END:VEVENT BEGIN:VEVENT SUMMARY:Taisiia Morozova (Uppsala University) DTSTART:20251105T121500Z DTEND:20251105T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/98 DESCRIPTION:Title: Multi-Agent Reinforcement Learning for Buffered Cellular Networks \nby Taisiia Morozova (Uppsala University) as part of Gothenburg stati stics seminar\n\nLecture held in MVL14.\n\nAbstract\nWe study the use of m ulti-agent reinforcement learning (MARL) for buffered cellular networks\, where base stations are modelled as independent agents making transmission decisions under interference and delay constraints. The network is descri bed through a stochastic geometry framework with Poisson-distributed base stations and users\, and buffers capturing traffic arrivals and service dy namics. To handle the interactions between agents\, we employ a mean-field approximation\, so that each agent responds to an aggregate distribution of its neighbours’ states and actions. The learning problem is formulate d via mean-field Q-learning\, where the objective is to improve network ca pacity while controlling delays. Initial experiments show convergence of t he Q-functions for several agents\, suggesting that the approach is well-s uited to this setting.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/98/ END:VEVENT BEGIN:VEVENT SUMMARY:Mathis Rost (Chalmers) DTSTART:20251022T111500Z DTEND:20251022T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/99 DESCRIPTION:Title: Void Probabilities and Likelihood Approximation for Gibbs Process es\nby Mathis Rost (Chalmers) as part of Gothenburg statistics seminar \n\nLecture held in MVL14.\n\nAbstract\nWhen fitting a model to data\, one would ideally like to use maximum likelihood estimation\, due to its nice statistical properties. Unfortunately\, the likelihood function\nof a gen eral Gibbs point process is typically not tractable\, due to the associate d normalizing constant. This has led to the development of a range of alte rnative methods\,\nsuch as Takacs-Fiksel estimation (including its special case pseudolikelihood estimation) and Point Process Learning.\nLeveraging recent probabilistic results for Gibbs processes\, in this talk we presen t an\napproach to perform approximate likelihood estimation for Gibbs proc esses. Specifically\, we show that the likelihood function can be expresse d completely in terms of\nthe Papangelou conditional intensity\, which is typically known and tractable. This\nnew likelihood representation involve s an infinite series expansion\, and we discuss\ndifferent ways of approxi mating it\, and thereby the likelihood function. We further\ndiscuss how t his plays out in certain models and compare it to the state-of-the-art.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/99/ END:VEVENT BEGIN:VEVENT SUMMARY:Simon Olsson (Chalmers) DTSTART:20260121T121500Z DTEND:20260121T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/100 DESCRIPTION:Title: Transferable Implicit Transfer Operators\nby Simon Olsson (C halmers) as part of Gothenburg statistics seminar\n\nLecture held in MVL14 .\n\nAbstract\nIn this talk\, I will outline our approach to use deep gene rative models to learn weak solutions to the Langevin equations with long time horizons. E.g. given an initial condition\, $x_0$\, learn the transit ion density $p_t(x_t\\mid x_0)$\, where $t$ is orders of magnitude larger than the usual numerical integration step. The context of this work is the $\\textit{sampling problem}$ from molecular dynamics\, an important metho d in chemistry\, physics\, and biology\, that faces slow mixing. I will gi ve numerous empirical examples of the successful application of this appro ach in molecular dynamics.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/100/ END:VEVENT BEGIN:VEVENT SUMMARY:Fabio Frommer (University of Mainz) DTSTART:20251210T121500Z DTEND:20251210T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/101 DESCRIPTION:Title: The inverse Henderson problem from statistical mechanics for mul ti-species models\nby Fabio Frommer (University of Mainz) as part of G othenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nThe in verse Henderson problem of statistical mechanics is the theoretical founda tion for many bottom-up coarse-graining techniques for the numerical simul ation of complex soft matter physics. This inverse problem concerns classi cal particles in continuous space interacting according to a pair potentia l depending on the distance of the particles. Roughly stated\, it asks for the interaction potential given the pair correlation function of the syst em. We show that the solution to this inverse problem is unique and can be rewritten as a minimization problem for a certain relative entropy funct ional. Lastly\, we show how this framework can be adapted to multi-specie s models using marked Gibbs measures.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/101/ END:VEVENT BEGIN:VEVENT SUMMARY:Zheng Zhao (Linköping University) DTSTART:20260218T121500Z DTEND:20260218T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/102 DESCRIPTION:Title: Diffusion differentiable resampling\nby Zheng Zhao (Linköpi ng University) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nThis work is concerned with differentiable resampling in the context of sequential Monte Carlo (e.g.\, particle filtering). We propose a new informative resampling method that is instantly pathwise dif ferentiable\, based on an ensemble score diffusion model. We prove that ou r diffusion resampling method provides a consistent estimate to the resamp ling distribution\, and we show by experiments that it outperforms the sta te-of-the-art differentiable resampling methods when used for stochastic f iltering and parameter estimation. Implementations are available online at https://github.com/zgbkdlm/diffres\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/102/ END:VEVENT BEGIN:VEVENT SUMMARY:Viktor Nilsson (KTH) DTSTART:20251028T100000Z DTEND:20251028T104500Z DTSTAMP:20260404T095849Z UID:gbgstats/103 DESCRIPTION:Title: New results in the large deviations of Schrödinger bridges\ nby Viktor Nilsson (KTH) as part of Gothenburg statistics seminar\n\nLectu re held in MVL14.\n\nAbstract\nIn a recent paper\, we show a large deviati on principle for certain sequences of static Schrödinger bridges\, typica lly motivated by a scale-parameter decreasing towards zero\, extending exi sting large deviation results to cover a wider range of reference processe s. Our results provide a theoretical foundation for studying convergence o f such Schrödinger bridges to their limiting optimal transport plans. Rec ently\, Bernton et al. established a large deviation principle\, in the sm all-noise limit\, for fixed-cost entropic optimal transport problems. In t his paper\, we address an open problem posed by Bernton et al. and extend their results to hold for Schrödinger bridges associated with certain seq uences of more general reference measures with enough regularity in a simi lar small-noise limit. These can be viewed as sequences of entropic optima l transport plans with non-fixed cost functions. Using a detailed analysis of the associated Skorokhod maps and transition densities\, we show that the new large deviation results cover Schrödinger bridges where the refer ence process is a reflected diffusion on bounded convex domains\, correspo nding to recently introduced model choices in the generative modeling lite rature.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/103/ END:VEVENT BEGIN:VEVENT SUMMARY:Martin Andrae (Linköping University) DTSTART:20260225T121500Z DTEND:20260225T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/105 DESCRIPTION:Title: Flow-based generative models for data assimilation\nby Marti n Andrae (Linköping University) as part of Gothenburg statistics seminar\ n\nLecture held in MVL14.\n\nAbstract\nFlow-based and diffusion generative models have emerged as powerful tools for sampling from complex\, high-di mensional distributions\, such as those found in image generation. In weat her forecasting\, they enable the generation of ensemble forecasts at a fr action of the computational cost of traditional numerical models. These mo dels have also shown promise for solving inverse problems like data assimi lation\, offering advantages over classical methods in high-dimensional\, nonlinear settings. In this talk\, I will introduce the core ideas behind these approaches and present some of our recent results.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/105/ END:VEVENT BEGIN:VEVENT SUMMARY:Filip Rydin (Chalmers\, E2) DTSTART:20260311T121500Z DTEND:20260311T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/106 DESCRIPTION:Title: Learning-based methods for vehicle routing problems - recent adv ances\nby Filip Rydin (Chalmers\, E2) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nThis talk reviews recent a dvances in machine learning for combinatorial optimization\, with a partic ular focus on routing problems such as the Traveling Salesman Problem (TSP ) and the Capacitated Vehicle Routing Problem (CVRP).\n\nFirst\, I will pr esent a unifying high-level hierarchy of methods. I will then delve deeper into end-to-end reinforcement learning approaches\, which have shown stro ng empirical performance. Finally\, I will present our recent work on mult i-objective routing over multigraphs\, highlighting how learning-based mod els can handle competing objectives and complex network structures.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/106/ END:VEVENT BEGIN:VEVENT SUMMARY:Jun Yang (Department of Mathematical Sciences\, University of Cope nhagen) DTSTART:20260129T121500Z DTEND:20260129T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/109 DESCRIPTION:Title: Stereographic Barker’s MCMC Proposal: Efficiency and Robustnes s at Your Disposal\nby Jun Yang (Department of Mathematical Sciences\, University of Copenhagen) as part of Gothenburg statistics seminar\n\nLec ture held in MVL14.\n\nAbstract\nWe introduce a new family of robust gradi ent-based MCMC samplers under the framework of stereographic MCMC (Yang et al. 2022) which maps the original high dimensional problem in Euclidean s pace onto a sphere. Compared with the existing Stereographic Projection Sa mpler (SPS) which is of a random-walk Metropolis type algorithm\, our new family of samplers is gradient-based using the Barker proposal (Livingston e and Zanella\, 2022)\, which improves SPS in high dimensions and is robus t to tuning. Meanwhile\, the proposed algorithms enjoy all the good proper ties of SPS\, such as uniform ergodicity for a large class of heavy and li ght-tailed distributions and "blessings of dimensionality".\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/109/ END:VEVENT BEGIN:VEVENT SUMMARY:Easter break DTSTART:20260408T111500Z DTEND:20260408T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/110 DESCRIPTION:by Easter break as part of Gothenburg statistics seminar\n\nLe cture held in MVL14.\nAbstract: TBA\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/110/ END:VEVENT BEGIN:VEVENT SUMMARY:Karl Hammar (Chalmers\, SAAB) DTSTART:20260423T111500Z DTEND:20260423T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/111 DESCRIPTION:Title: Low-Variance Importance Sampling for Discretely Observed Stochas tic Differential Equations\nby Karl Hammar (Chalmers\, SAAB) as part o f Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nSto chastic differential equations (SDEs) are commonly used to model dynamical systems of interest. When such systems are observed at discrete times\, t hey give rise to continuous–discrete state space models\, where the main tasks are Bayesian inference of the latent state (filtering and smoothing ) as well as computation of the observation likelihood (for model comparis on). In linear and Gaussian settings these problems can be solved efficien tly with recursive algorithms like the Kalman filter\, but in nonlinear ca ses Kalman-type solutions require approximations and lead to biased soluti ons.\n\nTo overcome these problems\, alternative solutions include sequent ial importance sampling (SIS)\, or perhaps most commonly\, sequential impo rtance resampling (SIR)\, also known as particle filters. These methods ar e unbiased and converge weakly to the correct solution as the number of pa rticles goes to infinity. However\, the efficiency of these methods depend s greatly on the choice of importance distribution\, as poor choices lead to high-variance importance weights and particle degeneracy. The design of good importance sampling distributions is therefore of crucial importance . For smoothing\, filtering\, and estimation of the observation likelihood \, the optimal importance distribution is given by the smoothing distribut ion\, which is the focus of this work.\n\nBy Doob’s h-transform\, the sm oothing distribution can be characterized as the law of a controlled SDE t hat differs from the unconditional one only by an additional drift term th at steers trajectories toward future observations. In this work\, we appro ximate this control term using neural networks\, yielding a tractable appr oximation of the smoothing distribution that can be corrected with low-var iance importance weights. The model is trained using divergence-based obje ctives\, including the Kullback–Leibler divergence\, and evaluated in te rms of effective sample size and variance of likelihood estimates. This ap proach reduces variance in importance sampling and improves the efficiency of inference in nonlinear SDE models.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/111/ END:VEVENT BEGIN:VEVENT SUMMARY:Karthik Bharath (University of Nottingham) DTSTART:20260415T111500Z DTEND:20260415T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/113 DESCRIPTION:Title: Rolled Gaussian process models for data as curves on manifolds\nby Karthik Bharath (University of Nottingham) as part of Gothenburg st atistics seminar\n\nLecture held in MVL14.\n\nAbstract\nGiven a planar cur ve\, imagine rolling a sphere along that curve without slipping or twistin g\, and by this means tracing out a curve on the sphere. Such a rolling op eration induces a local isometry between the sphere and the plane so that the two curves uniquely determine each other\, and moreover\, the operatio n extends to a general class of manifold $M$ in any dimension $d$. \n\nI w ill describe how rolling can be used to construct an analogue of a Gaussia n process with values in $M$\, known as a rolled Gaussian process\, starti ng from an $\\mathbb R^d$-valued Gaussian process with mean $m$ and covari ance $K$. I will discuss the relationship between $m$ and the Frechet mean of the rolled process\, and using the inverse operations of unrolling and unwrapping\, discuss simple estimators of $m$ and $K$ and their convergen ce rates. Utility of the model will be shown in an application involving c urves on 3D orientations coming from a robot learning experiment.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/113/ END:VEVENT BEGIN:VEVENT SUMMARY:Isac Boström (Chalmers) DTSTART:20260318T121500Z DTEND:20260318T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/114 DESCRIPTION:Title: Half-way seminar: Bayesian Inference for Models of Text Data \nby Isac Boström (Chalmers) as part of Gothenburg statistics seminar\n\n Lecture held in MVL14.\n\nAbstract\nModels of text data are increasingly a pplied to inference tasks in the social sciences to investigate a wide ran ge of linguistic and cultural phenomena. Word embeddings\, for example\, a re commonly used to study semantic change\, political language\, and socia l bias in large collections of text. However\, these models are typically estimated by optimization\, producing point estimates without principled u ncertainty quantification.\n\nIn this talk\, I present a Bayesian formulat ion of probabilistic word embedding models\, focusing on skip-gram with ne gative sampling and briefly discussing continuous bag-of-words. I explain why the posterior distribution is non-identifiable under general linear tr ansformations of the embedding space and introduce a simple and principled constraint that ensures a well-defined posterior. I then compare differen t approaches to posterior inference\, including mean-field variational inf erence\, Hamiltonian Monte Carlo\, and Pólya-Gamma Gibbs sampling. By aug menting the likelihood with Pólya-Gamma latent variables\, we obtain an e fficient sampler that provides scalable and well-calibrated uncertainty qu antification. \n\nI will also briefly discuss the structural topic model a s a related example where Bayesian uncertainty plays a central role.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/114/ END:VEVENT BEGIN:VEVENT SUMMARY:Clara Bertinelli Salucci DTSTART:20260429T111500Z DTEND:20260429T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/115 DESCRIPTION:by Clara Bertinelli Salucci as part of Gothenburg statistics s eminar\n\nLecture held in MVL14.\nAbstract: TBA\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/115/ END:VEVENT BEGIN:VEVENT SUMMARY:Shuangshuang Chen (KTH) DTSTART:20260506T111500Z DTEND:20260506T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/116 DESCRIPTION:by Shuangshuang Chen (KTH) as part of Gothenburg statistics se minar\n\nLecture held in MVL14.\nAbstract: TBA\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/116/ END:VEVENT BEGIN:VEVENT SUMMARY:Fredrik Sävje (Uppsala University\, Department of Economics) DTSTART:20261028T121500Z DTEND:20261028T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/118 DESCRIPTION:by Fredrik Sävje (Uppsala University\, Department of Economic s) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\nAbs tract: TBA\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/118/ END:VEVENT BEGIN:VEVENT SUMMARY:Måns Magnusson (Uppsala University\, Department of Statistics) DTSTART:20261104T121500Z DTEND:20261104T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/119 DESCRIPTION:by Måns Magnusson (Uppsala University\, Department of Statist ics) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\nA bstract: TBA\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/119/ END:VEVENT BEGIN:VEVENT SUMMARY:Andrea Zanoni (Scuola Normale Superiore) DTSTART:20260528T111500Z DTEND:20260528T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/121 DESCRIPTION:Title: Learning interaction kernels in stochastic particle systems\ nby Andrea Zanoni (Scuola Normale Superiore) as part of Gothenburg statist ics seminar\n\nLecture held in MVL14.\n\nAbstract\nInference in stochastic interacting particle systems is increasingly important due to application s in social sciences\, physics\, and machine learning. In this talk\, we f ocus on learning the interaction kernel from observations of a single part icle. We adopt a semi-parametric approach\, expressing the kernel as a gen eralized Fourier series with orthogonal polynomials tailored to the proble m. The Fourier coefficients are estimated via a variation of the method of moments applied to the invariant measure of the mean-field dynamics\, res ulting in a linear system based on moments approximated from the particle trajectory. We analyze the approximation error and asymptotic behavior of the estimator in the limits of infinite observation time\, large particle number\, and increasing number of Fourier coefficients. Numerical experime nts illustrate the effectiveness of the approach. This work is joint with Grigorios A. Pavliotis (Imperial College London).\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/121/ END:VEVENT BEGIN:VEVENT SUMMARY:Piotr Graczyk (LAREMA Université d'Angers) DTSTART:20260527T111500Z DTEND:20260527T120000Z DTSTAMP:20260404T095849Z UID:gbgstats/122 DESCRIPTION:Title: Penalized estimation for Big Data in Regression Problems and its Geometry\nby Piotr Graczyk (LAREMA Université d'Angers) as part of G othenburg statistics seminar\n\nLecture held in MVL14.\n\nAbstract\nI will present recent results obtained in [1] and [2] jointly with M. Bogdan\, X . Dupuis\, B. Kolodziejek\, U. Schneider\, T. Skalski\, P. Tardivel and M. Wilczynski.\n\nIt is well known that LASSO discovers zero coefficients of the vector $b$ in the\nregression equation $Y=Xb+\\varepsilon$ where $X$ is the data matrix and $Y$ the response vector. In fact LASSO estimates th e sign of the coefficient vector $b$ ($ b_i$'s positive\, negative or nul l). The sign is called the model(pattern) of LASSO. In the LASSO estimator the\n$\\ell^1$ penalty is employed.\n\nIn the study of Big Data one needs to identify more informative patterns of the vector $b$. These leads to use penalties different from the $\\ell^1$ penalty and to get more dime nsionality reduction. \n\nWe define the pattern of any estimator with poly hedral penalty\, i.e. the unit ball $B$ with respect to the penalty norm is a convex polyhedron. Surprising links between the pattern of a penalize d estimator and the geometry of the convex polytope $B^*$ will be explaine d.\n\n\n\n We study in detail estimation with a sorted $\\ell^1$ penalty\ , called SLOPE. Its dual ball $B^*$\n is a signed permutahedron. \n SLOP E is a popular method for dimensionality reduction in the high-dimensional regression\, encompassing the LASSO estimator but also the $l^\\infty$ p enality. Indeed\, some coefficients of the estimator $\\hat b ^{\\rm SL OPE}$ are null (sparsity) and others are equal in absolute value (cluster ing). Consequently\, irrelevant predictors are eliminated and groups of predictors having the same influence on the\nresponse vector are identifi ed.\nThe SLOPE pattern of a vector $b$ provides: the sign of its component s\, clusters (components equal in absolute value) and clusters ranking.\n \n In our research we give an analytical necessary and sufficient conditio n for SLOPE pattern recovery of an unknown vector $b$ of regression coeffi cients. Such condition is called Irrepresentability(IR) condition. For any polyhedral penalty we find a geometric IR condition.\n\n[1] P. Graczyk\, U. Schneider\, T. Skalski\, P. Tardivel\, A Unified Framework for Pattern Recovery in Penalized and Thresholded Estimation and its Geometry\, Jour nal of Optimization Theory and Applications(2026) vol. 208(1)\, 1-41.\n\n[ 2] M. Bogdan\, X. Dupuis\, P. Graczyk\, B. Kolodziejek\, T. Skalski\, P. Tardivel\,\nM. Wilczynski\, Pattern recovery by SLOPE\, \nApplied and Com putational Harmonic Analysis 80(2026)\, 1-25.\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/122/ END:VEVENT BEGIN:VEVENT SUMMARY:Peter Rudzis (Chalmers) DTSTART:20260325T121500Z DTEND:20260325T130000Z DTSTAMP:20260404T095849Z UID:gbgstats/124 DESCRIPTION:Title: Brownian particle systems with singular interactions\nby Pet er Rudzis (Chalmers) as part of Gothenburg statistics seminar\n\nLecture h eld in MVH12.\n\nAbstract\nWe present a class of Brownian interacting part icle systems known as \\textit{rank-based diffusions} and their alter ego\ , \\textit{systems of competing Brownian particles}. The former originally appeared as a model in stochastic portfolio theory\, while the latter mod el—obtained by considering the order statistics of the former—is relat ed to skew-reflected Brownian motion. This talk will be mainly expository\ , describing for a broad audience the fundamental properties of these proc esses\, including their associated stationary distributions. We will also discuss the infinite-particle versions of these models\, where the station arity structure is richer. As a representative calculation\, we will show that the distribution of the lowest particle in equilibrium is often Gumbe l or related. Finally\, we will describe some of our results on the equili brium fluctuations of a certain space-time random field associated with th e infinite Atlas model (a prototypical model in the class of rank-based di ffusions). These fluctuations have a scaling limit given by a two-paramete r Gaussian process with explicit covariance structure\, equivalently descr ibed as the solution to a certain stochastic partial differential equation (SPDE). As a result\, tagged particles exhibit fluctuations that locally behave as fractional Brownian motion with Hurst parameter 1/4. This work i s joint with Sayan Banerjee and Amarjit Budhiraja (UNC Chapel Hill).\n LOCATION:https://stable.researchseminars.org/talk/gbgstats/124/ END:VEVENT END:VCALENDAR