Date: 29^th December 2024, 09:00 to 12:30

Title: Omics Data Science Workshop

Instructor: Himel Mallick, Cornell University

Description: Due to the vast diversity of available technologies, there is a dire need to impose some order onto the current inventory of omics data science algorithms, which are mushrooming at a staggering rate, to better understand their statistical core. The goal of this workshop is to teach fundamental principles that unify our understanding of seemingly unrelated and strikingly diverse data streams in omics sciences. By simplifying and unifying approaches that are common across omics, this course aims to decrease entry barriers into omics data science while remaining cognizant of omics-specific nuances and complexities in a coherent manner.

While there are short courses focusing on platform-specific tools and methods, researchers and practitioners would greatly benefit from a dedicated short course on unified omics data science. This workshop will provide an overview of the field of omics data science from a holistic standpoint, with a focus on the platform-agnostic tools that are used to analyze vastly different kinds of omics data, including genomics, metagenomics, transcriptomics, metabolomics, and proteomics, thus enhancing the interoperability of disparate omics studies. We will cover topics such as normalization, differential analysis, integration, machine learning, and visualization.

The workshop will be taught by Dr. Himel Mallick, from Cornell University. Dr. Mallick is the developer of several state-of-the-art tools that will be covered in this workshop, such as MaAsLin2, Tweedieverse, and IntegratedLearner. The workshop is geared towards researchers and practitioners interested in honing their omics data science skillset. Familiarity with basic biology and R/Bioconductor is a plus but not required.

Date: 30^th December 2024, 09:00 to 12:30

Title: Interpretable and Interactive Machine Learning

Instructor: Kris Shankaran, University of Wisconsin-Madison

Description: Our society’s capacity for algorithmic problem-solving has perhaps never been higher — the combination of powerful abstractions, plentiful data, and accessible software means that machine learning is now a central tool across various industries and areas of research. At the same time, we are regularly warned of the risks of misusing this capacity (e.g., the New York Times in 2024: “Imran Khan’s ‘Victory Speech’ From Jail Shows A.I.’s Peril and Promise,” “The Trouble With A.I. Sharpening,” “Are A.I. Mammograms Worth Their Cost?”). Interpretable and interactive machine learning aims to make complex models understandable and controllable, enhancing user agency and creativity. This course will review principles from the growing literature on this topic. We will introduce precise vocabulary for discussing interpretability, for example, the distinction between glass-box and explainable algorithms. Connections to classical statistical and design principles, like parsimony and the gulfs of interaction, will be highlighted. We will review the mechanics of basic explainability techniques, including feature embeddings, integrated gradients, and concept bottlenecks. Real-world applications and simplified code demos will illustrate how interpretability can support progress on longstanding challenges, like out-of-domain generalization and fairness. Criteria for objectively evaluating interpretability mechanisms will be discussed. The role of the audience’s perspective on interactive algorithm design will be emphasized. Finally, we will explore open problems and the potential role of statistical thinking in addressing them.