Statistical Learning & Inference Seminars
The seminars will take place every Tuesday 11am-12pm during term time.
Term 1, 25-26
|
Date,Time and Room |
Speaker |
Title |
|
|
07/10, 11am MB2.22 |
Vasiliki Koutra (King's College London)Link opens in a new window |
Designing experiments on networks |
|
|
Abstract: |
The design of experiments provides a formal framework for the collection of data to aid decision making. When such experiments are performed on connected units linked through a network, the resulting design and analysis are more complex; e.g. is the observed response from a given unit due to the direct effect of the treatment applied to that unit or the result of a network, or viral, effect arising from treatments applied to connected units? In this talk, I propose a methodology for constructing efficient designs which control for variation among the experimental units arising from network interference, so that the direct treatment effects can be precisely estimated. Performance gains over conventional designs will be demonstrated via different example experiments. |
||
|
14/10, 11am MB2.22 |
Johan Segers (KU Leuven)Link opens in a new window | Statistics for numerics: improving Monte Carlo integration by control variates | |
|
Abstract: |
Numerical integration of a function is a ubiquitous problem in applied mathematics, with applications in statistics, machine learning, finance, ... Monte Carlo methods have the advantage that they require very little regularity and provide a convergence rate that does not depend on the dimension of the integration domain. The use of control variates is one out of many methods to reduce the variance of the Monte Carlo estimate of the integral. In its simplest form, the control variate method can be cast as a multiple linear regression problem. Letting the number of control variates tend to infinity with the number of Monte Carlo particles can yield faster convergence rates but raises the problem of variable selection, which can be tackled by the lasso. Control variates can also be adapted to more sophisticated Monte Carlo schemes such as adaptive importance sampling. Finally, nearest neighbour estimates can act as control variates to speed up the convergence rate of the Monte Carlo procedure by an amount that depends on the intrinsic dimension of the domain and the regularity of the integrand. The talk is based on joint work with Rémi Leluc, Aymeric Dieuleveut, François Portier, and Aigerim Zhuman. |
||
|
21/10, 11am MB2.22 |
Nikolas Nusken (King's College London)Link opens in a new window | Go with the flow | |
|
Abstract: |
Fix a curve of probability distributions, perhaps connecting prior and posterior in Bayesian inference. In this talk, I will survey recent ideas to construct interacting particle systems, designed to follow this flow as closely as possible, and hence solve the Bayesian inference problem. We will touch on Stein variational gradient descent, kernel mean embeddings, optimal control and, perhaps, diffusion models. | ||
|
28/10, 11am MB2.22 |
Francois-Xavier Briol (UCL)Link opens in a new window | Multilevel neural simulation-based inference | |
|
Abstract: |
Neural simulation-based inference (SBI) is a popular set of methods for Bayesian inference when models are only available in the form of a simulator. These methods are widely used in the sciences and engineering, where writing down a likelihood can be significantly more challenging than constructing a simulator. However, the performance of neural SBI can suffer when simulators are computationally expensive, thereby limiting the number of simulations that can be performed. In this paper, we propose a novel approach to neural SBI which leverages multilevel Monte Carlo techniques for settings where several simulators of varying cost and fidelity are available. We demonstrate through both theoretical analysis and extensive experiments that our method can significantly enhance the accuracy of SBI methods given a fixed computational budget. |
||
|
04/11, 11am MB2.22 |
Mona Azadkia (LSE)Link opens in a new window | Coverage correlation: detecting singular dependencies between random variables | |
|
Abstract: |
We introduce the coverage correlation coefficient, a novel nonparametric measure of statistical association that quantifies the extent to which two random variables have a joint distribution concentrated on a singular subset relative to the product of their marginals. Our statistic consistently estimates an f-divergence between the joint distribution and the product of the marginals—taking the value 0 if and only if the variables are independent, and 1 if and only if the copula is singular. By leveraging Monge–Kantorovich ranks, the coverage correlation naturally extends to multivariate settings, allowing it to capture associations between random vectors. The method is distribution-free, has an analytically tractable asymptotic null distribution, and is computationally efficient, making it well-suited for detecting complex, potentially nonlinear dependencies in large-scale pairwise testing. | ||
|
11/11, 11am MB2.22 |
Matteo Barigozzi (Università di Bologna)Link opens in a new window | Estimation of large approximate dynamic matrix factor models based on the EM algorithm and Kalman filtering | |
|
Abstract: |
This paper considers an approximate dynamic matrix factor model that accounts for the time series nature of the data by explicitly modelling the time evolution of the factors. We study estimation of the model parameters based on the Expectation Maximization (EM) algorithm, implemented jointly with the Kalman smoother which gives estimates of the factors. We establish the consistency of the estimated loadings and factor matrices as the sample size T and the matrix dimensions p1 and p2 diverge to infinity. We then illustrate two immediate extensions of this approach to: (a) the case of arbitrary patterns of missing data and (b) the presence of common stochastic trends. The finite sample properties of the estimators are assessed through a large simulation study and two applications on: (i) a financial dataset of volatility proxies and (ii) a macroeconomic dataset covering the main euro area countries. | ||
|
18/11, 11am MB2.22 |
Myrto Limnios (EPFL)Link opens in a new window | Testing for Causal Effects based on Event Processes | |
|
Abstract: |
In the context of disease progression analysis, estimating the causal effect of a time-continuous treatment assigned to a given population is an important problem. This is also relevant in applications that gather massive high-dimensional and time-dependent data structures for understanding causal relationships, e.g., in financial and social studies.
In this work, we propose a nonparametric model for event processes, based on a linear combination of tensor products of kernel functions, composed of stochastic integrals w.r.t. the event processes observed up to a fixed time. Under some assumptions of sparse decomposition and adequate regularity, the optimal parameters involved in the expansion minimize a lasso risk. Finite-sample guarantees for the estimation and prediction errors are proved, that yield data-driven optimal weights for the LASSO penalty term allowing for heteroscedastic expansions.
The results will be applied to propose a conditional local independence test, that is fundamental to learning causal graphs. This is a joint work with Prof. Niels R. Hansen (Copenhagen University, Denmark).
|
||
|
25/11, 11am MB2.22 |
Timothy Cannings (Edinburgh)Link opens in a new window | Nonparametric classification with missing data | |
|
Abstract: |
We introduce a new nonparametric framework for classification problems in the presence of missing data. The key aspect of our framework is that the regression function decomposes into an anova-type sum of orthogonal functions, of which some (or even many) may be zero. Working under a general missingness setting, which allows features to be missing not at random, our main goal is to derive the minimax rate for the excess risk in this problem. In addition to the decomposition property, the rate depends on parameters that control the tail behaviour of the marginal feature distributions, the smoothness of the regression function and a margin condition. The ambient data dimension does not appear in the minimax rate, which can therefore be faster than in the classical nonparametric setting. We further propose a new method, called the Hard-thresholding Anova Missing data (HAM) classifier, based on a careful combination of a k-nearest neighbour algorithm and a thresholding step. The HAM classifier attains the minimax rate up to polylogarithmic factors and numerical experiments further illustrate its utility. | ||
|
02/12, 11am MB2.22 |
Reza Drikvandi (Durham)Link opens in a new window | ||
|
Abstract: |
|||
|
09/12, 11am MB2.22 |
Takeru Matsuda (University of Tokyo)Link opens in a new window | ||
|
Abstract: |
|||