Presenter: Milena Pavlovic
Title
“Formalization of machine learning problems for life sciences: insights from the adaptive immune receptor field”.
Abstract
Despite the increasingly prevalent use of machine learning for life science research, reproducibility, comparability, and generalization of results remain limited. We argue that this gap reflects the implicit and inconsistent translation of biological questions into machine learning problems, leaving unexamined the key assumptions about the data-generating process, the quantity being estimated, and the conditions under which conclusions hold. We propose a reasoning framework to formalize this process and reflect on the correspondence of the experimental setup to the underlying property of interest. We use adaptive immune receptors to demonstrate how the framework applies across different machine learning tasks and how implicit assumptions lead to conclusions that are difficult to interpret or compare.
BiLS chair
Jonas Paulsen