Overview

Description

CALENDAR DESCRIPTION:

Current topics in artificial intelligence depending on faculty and student interest.

COURSE DETAILS:

This is a research-oriented course on two machine learning topics that are central to the Ester lab, transfer learning and machine learning for precision medicine. The instructor will first introduce these topics through tutorials. Students will then choose a recent paper on one of these topics and present it to the class. Finally, students will define a course research project, usually related to the chosen paper, present a project proposal to the class and submit a project report. Ideally, the course project will, possibly after some extra work, lead to a publication.

Transfer learning

In many cases, one does not have enough data from your target domain to train a machine learning model. Simply applying a model trained on a source domain with more data to the target domain does not work, if the IID (independent and identically distributed) assumption is violated. Transfer learning relaxes the assumption for various scenarios, such as domain adaptation (source domain with lots of labeled data, target domain with only unlabeled data), domain generalization (several labelled source domains, no target domain data), and continual learning (series of domains with labelled data). In order to deal with the distribution shift between different domains, many transfer learning methods learn a latent representation space in which the different domains have a similar distribution. Another challenge of transfer learning is avoiding catastrophic forgetting, and we will discuss regularization and replay-based approaches.

ML for precision medicine

The goal of precision medicine is to diagnose and treat patients more accurately and effectively by considering the individual molecular profile of the patient, rather than using a "one-size-fits-all" approach based only on the symptoms. Precision medicine is based on rich and complex patient data, in particular mutation data, gene expression data, etc. The complexity of the data makes it impossible to analyze it manually, motivating the need for machine learning methods. Machine learning models can, e.g., perform diagnosis, prognosis, treatment recommendation, drug target identification, and drug discovery. Specific challenges of machine learning for precision medicine include the small size of many datasets, suggesting the need for transfer learning, and the requirement that predictions are explainable, e.g. through causal models of mechanisms.

COURSE-LEVEL EDUCATIONAL GOALS:

Transfer learning

• Introduction
• Domain adaptation
• Domain Generalization
• Continual learning

ML for precision medicine

• Introduction
• Drug response prediction
• Discover of genetic causes of adverse drug reactions
• Drug discovery (Generative molecule models)

Grading

Materials

REQUIRED READING NOTES: