Score-based diffusion models have emerged as a powerful framework for generative modeling, progressively transforming noise into structured data samples when access to a dataset is available. In this talk, we explore how to extend these ideas to the sampling setting, where the target distribution is only known up to a normalizing constant. After reviewing the fundamentals of diffusion models, we highlight a key observation: the score function central to these methods can be expressed as an expectation with respect to a time-dependent distribution with known unnormalized density. This perspective motivates Stochastic Localization via Iterative Posterior Sampling (SLIPS), an approach that estimates the score function using Monte Carlo methods and leverages it to construct a denoising process. We will examine the theoretical underpinnings of SLIPS, with particular emphasis on its main limitation, the duality of log-concavity, which restricts its practical applicability. Building on this, I will present a new approach to Iterative Posterior Sampling (forthcoming work) that bypasses explicit score estimation altogether, leading to significantly improved scalability. While this method remains affected by the same duality phenomenon, we will see that its impact is mitigated in practice.

To join this seminar virtually, please request Zoom connection details from hr.ops@stat.ubc.ca.

The van Eeden seminar is a yearly event in which graduate students vote for their favorite statisticians. The winner is contacted by the organizing committee and invited to give a talk in the department’s seminar. The speaker spends one or two days on-campus, and graduate students have the opportunity to have lunch and dinner with them.

THIS YEAR'S SPEAKER

The Constance van Eeden Speaker for 2026 is Dr. Ryan Tibshirani, Professor in the Department of Statistics at the University of California, Berkeley, and Principal Investigator in the Delphi Research Group. Before joining Berkeley, Dr. Tibshirani served as a faculty member in the Departments of Statistics and Machine Learning at Carnegie Mellon University from 2011 to 2022. He earned his Ph.D. in Statistics from Stanford University in 2011 under the supervision of Professor Jonathan Taylor, and his B.S. in Mathematics from Stanford University in 2007. 

Seminar Title: Online Conformal Prediction, Multi-Level Quantile Tracking, and Gradient Equilibrium

Event Date: Thursday, April 2nd, 2026. 10:30-12:00.
Location: ESB 5104, University of British Columbia*

Event registration: https://ubc.zoom.us/meeting/register/htGYWnrFSCWhHIs8-d4wqw

Abstract:

This talk is about uncertainty quantification for time series prediction.

The overarching goal is to provide easy-to-use algorithms with formal guarantees. The algorithms we present build upon ideas from conformal prediction and control theory, are able to prospectively model conformal scores in an online setting, and adapt to the presence of systematic errors due to seasonality, trends, and general distribution shifts. We will then discuss an extension of these ideas to the setting of probabilistic forecasting, which is essentially a generalization of the framework to handle vector-valued predictions, i.e., predictions which take the form of a set of ordered quantile forecasts at different probability levels. Finally, we will generalize this even further to discuss an abstract property in online learning called gradient equilibrium, which encapsulates these settings, and more.

Dr. Ryan Tibshirani has been invited to be this year’s van Eeden speaker by the graduate students in the Department of Statistics at the University of British Columbia. A van Eeden speaker is a prominent statistician who is chosen each year to give a lecture, supported by the UBC Constance van Eeden Fund. The 2024 seminar is additionally sponsored by the Canadian Statistical Sciences Institute (CANSSI), the Pacific Institute for the Mathematical Sciences (PIMS), and the Walter H. Gage Memorial Fund.

*The room location may change.

Weighting methods are essential tools for estimating causal effects in observational studies, with the goal of balancing pre-treatment covariates across treatment groups. Traditional approaches pursue this objective indirectly, for example, via inverse propensity score weighting or by matching a finite number of covariate moments, and therefore do not guarantee balance of the full joint covariate distributions. Recently, distributional balancing methods have emerged as robust, nonparametric alternatives that directly target alignment of entire covariate distributions, but they lack a unified framework, formal theoretical guarantees, and valid inferential procedures. We introduce a unified framework for nonparametric distributional balancing based on the characteristic function distance (CFD) and show that widely used discrepancy measures, including the maximum mean discrepancy and energy distance, arise as special cases. Our theoretical analysis establishes conditions under which the resulting CFD-based weighting estimator achieves root-N consistency. Since the standard bootstrap may fail for this estimator, we propose subsampling as a valid alternative for inference. We further extend our approach to an instrumental variable setting to address potential unmeasured confounding. Finally, we evaluate the performance of our method through simulation studies and a real-world application, where the proposed estimator performs well and exhibits results consistent with our theoretical predictions.

The paper is available at https://arxiv.org/abs/2601.15449

Bio:

Dr. Chan Park is an assistant professor at the University of Illinois Urbana-Champaign. His research focuses on causal inference in complex settings, including dependence among units and omitted variables. He specializes in applying nonparametric methods and semiparametric theory to address these challenges.

To join this seminar virtually, please request Zoom connection details from hr.ops@stat.ubc.ca.

In practical regression applications, multiple covariates are often measured, but not all may be associated with the response variable. Identifying and including only the relevant covariates in the model is crucial for improving prediction accuracy. In this work, we develop a variational inference approach for estimation and variable selection in scalar-on-function regression, involving only functional covariates, and in partially functional regression models that also include scalar covariates. Specifically, we develop a variational expectation–maximization algorithm, with a variational Bayes procedure implemented in the E-step to obtain approximate marginal posterior distributions for most model parameters, except for the regularization parameters, which are updated in the M-step. Our method accurately identifies relevant covariates while maintaining strong predictive performance, as demonstrated through extensive simulation studies across diverse scenarios. Compared with alternative approaches, including BGLSS (Bayesian Group Lasso with Spike-and-Slab priors), grLASSO (group Least Absolute Shrinkage and Selection Operator), grMCP (group Minimax Concave Penalty), and grSCAD (group Smoothly Clipped Absolute Deviation), our approach achieves a superior balance between goodness-of-fit and sparsity in most scenarios. We further illustrate its practical utility through real-data applications involving spectral analysis of sugar samples and weather measurements from Japan.
 

To join this seminar virtually, please request Zoom connection details from hr.ops@stat.ubc.ca. 

Machine learning methods are now increasingly studied and used in National Statistical Offices, in particular to handle item nonresponse, where some survey respondents answer certain questions but leave others missing. In most surveys, item nonresponse affects key study variables, and imputation is routinely used to handle the resulting missing data. Standard parametric imputation methods can support rigorous inference when their modeling assumptions are approximately correct. However, when the imputation model is misspecified, the resulting inferences may be potentially misleading. Machine learning offers a flexible alternative by learning complex relationships between variables from the data, which can reduce the risk of misspecification. At the same time, this flexibility introduces new challenges for survey inference, since modern learning algorithms may converge more slowly than classical parametric models and may not automatically deliver valid uncertainty quantification. In this talk, I will present a survey sampling extension of the double/debiased machine learning framework of Chernozhukov et al. (2018). The proposed approach combines machine learning-based imputation with design-based survey weighting and an orthogonalized estimating strategy, leading to root-$n$ consistent and asymptotically normal estimation of population means under realistic conditions. We also develop a consistent variance estimator, yielding asymptotically valid confidence intervals while allowing the use of a wide range of machine learning algorithms. I will briefly discuss aggregation procedures and conclude with simulation results illustrating the performance of the proposed methodology.
 

This talk is part of the UBC Statistics Colloquium Series, which features broad and accessible seminars throughout the term and is sponsored in part by the Constance van Eeden Endowment.

In this talk, we propose a state space model for functional time series data, which extends many time series models to the realm of functional data. Most notably, we introduce the Functional ARMAX process (FARMAX), which is developed in the fully functional setting, i.e. without relying on projection onto a finite number of basis functions. These models are fit via our fully functional variant of the Kalman filter and smoother methods. The theoretical soundness of this approach is proven using tools from the theory of Gaussian measures in locally convex spaces. As an application, we consider signals data collected from small wearable medical tri-axial accelerometers affixed to a patient's wrists or ankles. Each device collects three time series (x, y, z directions) at 100Hz and can continuously collect data for 14 days.

//--Note updated time is 11 AM, March 10th--//

To join this seminar virtually, please request Zoom connection details from hr.ops@stat.ubc.ca. 

To design Bayesian studies, criteria for the operating characteristics of posterior analyses—such as power and the Type I error rate—are often assessed by estimating sampling distributions of posterior probabilities via simulation. In this work, we propose an economical method to determine optimal sample sizes and decision for such studies. Using our theoretical results that model posterior probabilities as a function of the sample size, we assess operating characteristics throughout the sample size space given simulations conducted at only two sample sizes. These theoretical results are used to construct bootstrap confidence intervals for the sample sizes and decision criteria that reflect the stochastic nature of simulation-based design. The broad applicability and wide impact of our methodology is illustrated using two clinical examples.

To join this seminar virtually, please request Zoom connection details from ea@stat.ubc.ca. 

In computer experiments, Gaussian process (GP) models are widely employed for emulation. However, when both qualitative and quantitative factors are involved, especially when qualitative factors have many categories, GP-based emulation becomes challenging, and existing methods can become unwieldy due to the curse of dimensionality. Motivated by computer experiments for the design of a cooling system, we introduce a new tree-based GP model for emulating computer codes with high-cardinality qualitative factors, referred to as the category tree GP (ctGP). The proposed approach incorporates a tree structure to partition the categories of the qualitative factors, after which GP or mixed-input GP models are fitted to the simulation outputs within the leaf nodes. The splitting rule is designed to reflect the cross-correlations among the categories of the qualitative factors, which a recent theoretical study has identified as a key component for improving prediction accuracy, and a pruning procedure based on cross-validation error is introduced to further ensure strong predictive performance. An application to the design of a cooling system demonstrates that the proposed method not only yields substantial computational gains and accurate predictions, but also offers meaningful insights into the system by uncovering an interpretable tree structure. Furthermore, in this cooling system design problem, the computer code is capable of generating multiple responses in addition to a single objective response; to accommodate this, we extend the ctGP framework to handle multiple responses by introducing an additional categorical variable that indicates which response is associated with each experimental point. Finally, we complete the cooling system design study by addressing the corresponding global optimization problem using Bayesian optimization with ctGP and an expected-improvement-type criterion.

To join this seminar virtually, please request Zoom connection details from ea@stat.ubc.ca. 

Bio: Ray-Bing Chen is a Professor in the Institute of Statistics and Data Science at National Tsing Hua University. He received his Ph.D. in Statistics from the University of California, Los Angeles in 2003. Prof. Chen’s research interests include statistical and machine learning, statistical modeling, computer experiments, and optimal design. His work has been published in leading journals such as the Annals of Applied Statistics, Journal of Computational and Graphical Statistics, Statistics and Computing, Technometrics, Journal of Quality Technology and Computational Statistics and Data Science. In recognition of his contributions to the field, he was elected as an Elected Member of the International Statistical Institute in 2020.

Bayesian hypothesis testing using the Bayes factor is an alternative to hypothesis testing based on the p-Value. They are especially useful if one considers the p-postulate, which suggests that equal p-values, irrespective of sample size, should represent equal evidence against a null hypothesis, false. Bayes factors can, however, be computationally intensive and require a prior distribution. We define an extension of Jeffrey's approximate objective Bayes factor (eJAB) based on a generalization of the unit information prior. Its computation requires nothing more than the p-value and the sample size and it provides a measure of evidence that allows one to interpret the p-value in light of the associated sample size through the lens of an approximate Bayes factor corresponding to an objective prior. We apply eJAB to reexamine the evidence from 71,130 clinical trial findings with particular attention to contradictions between Bayes factors and NHST—i.e., instances of the Jeffreys–Lindley paradox (JLP). Our findings reflect increasing evidence in the literature of problematic clinical trial design and results.

To join this seminar virtually, please request Zoom connection details from ea@stat.ubc.ca.