You are here

Home » Research » Seminars & Colloquia

Statistics Seminar

The Statistics Seminar has talks on a variety of topics. For more information contact Yongli Sang.

Spring 2026

During the Spring 2026 semester we will meet on Friday from 1:00-2:00 in Maxim Doucet Hall room 212. On some occasions we will meet on zoom

For more information contact Yongli Sang.

  1. 20 February 2026 (On zoom. Please contact Yongli Sang to request the zoom information.)
    An Association Test Based on Kernel-Based Neural Networks for Complex Genetic Association Analysis
    Tingting Hou
    Department of Experimental Statistics
    LSU
    Abstract: The advent of artificial intelligence, especially the progress of deep neural networks, is expected to revolutionize genetic research and offer unprecedented potential to decode the complex relationships between genetic variants and disease phenotypes, which could mark a significant step toward improving our understanding of the disease etiology. While deep neural networks hold great promise for genetic association analysis, limited research has been focused on developing neural network-based tests to dissect complex genotype-phenotype associations. This complexity arises from the opaque nature of neural networks and the absence of defined limiting distributions. We have previously developed a kernel-based neural network model (KNN) that synergizes the strengths of linear mixed models (LMM) with conventional neural networks. KNN adopts a computationally efficient minimum norm quadratic unbiased estimator (MINQUE) algorithm and uses kernel-based neural network structure to capture the complex relationship between large-scale sequencing data and a disease phenotype of interest. In the KNN framework, we introduce a MINQUE-based test to assess the joint association of genetic variants with the phenotype, which considers non-linear and non-additive effects and follows a mixture of chi-square distributions. We also construct two additional tests to evaluate and interpret linear and non-linear/non-additive genetic effects, including interaction effects. Our simulations show that our method consistently controls the type I error rate under various conditions and achieves greater power than a commonly used sequence kernel association test (SKAT), especially when involving non-linear and interaction effects. When applied to real genetic data from the UK Biobank, our approach identified genes associated with hippocampal volume, which can be further replicated and evaluated for their role in the pathogenesis of Alzheimer's disease.
    About the Speaker:Tingting Hou is an Assistant Professor in the Department of Experimental Statistics at Louisiana State University. He received his Ph.D. in Biostatistics from the University of Florida. His research focuses on developing statistical and machine-learning methods for genomics, genetics, and high-dimensional biomedical data.
  2. 6 March 2026
    Matched-Pair Benchmarks: Measuring What ML Models Actually Know
    Leizhen Zhang
    School of Computing and Informatics
    UL Lafayette
    Abstract: AI models can score highly on standard benchmarks while relying on superficial cues rather than true reasoning or semantic understanding. This talk introduces matched-pair benchmarks, where test instances are constructed in near-identical pairs that differ only in a controlled underlying property that flips the correct answer. We apply this approach to (1) SAT-style logic problems evaluated across several large language models and (2) software vulnerability detection evaluated across multiple detectors. We propose Accurate Differentiation Ratio (ADR)—the fraction of pairs a model correctly distinguishes—as a complementary metric to precision/recall/F1 that better exposes shortcut-driven performance. We conclude with limitations of pair-based evaluation and a stronger extension to cluster-based benchmarks for robustness.