Few-Shot Learning Tutorial - With a Focus on Drug Discovery

Speaker: Johannes Schimunek, University of Cambridge
Series: Machine Learning Reading Group @ CUED
Also open online: the joining details are on the talk's page.

Predictive machine learning models have proven to be valuable tools across a wide range of applications. However, many of these models require large amounts of labeled training data to achieve good predictive performance. In real-world applications, such data is often scarce, limiting the applicability of conventional machine learning approaches. Few-shot learning addresses this challenge by developing methods specifically designed to learn from only a small number of labeled examples.

Drug discovery is a particularly relevant application, as experimental measurements are often expensive and time-consuming, leaving many drug discovery projects with only limited amounts of labeled data.

In this tutorial, we introduce the fundamental concepts of few-shot learning, using drug discovery as a running example. We distinguish between inductive and transductive few-shot learning and discuss representative methods from both settings. Within inductive few-shot learning, we cover gradient-based and non-gradient-based approaches, illustrating how different methods leverage prior knowledge to adapt to new prediction tasks with only a few labeled examples. Finally, we discuss practical challenges and limitations of few-shot learning, particularly in the context of drug discovery.