Meta-Learning and Generalization Across Tasks

A major limitation of deep learning in many real-world applications is the scarcity of labelled data. Meta-learning, which leverages experience across related tasks to accelerate learning on new problems, has the potential to mitigate this data scarcity. However, most existing meta-learning methods are developed and evaluated under idealised assumptions, such as homogeneous task definitions, fixed data modalities, and well-defined training distributions.

In this research area, we study meta-learning as a foundation for building robust and adaptable machine-learning systems that transfer readily across tasks with only few training data points. Specifically, we focus on meta-learning methods that operate under realistic conditions, including heterogeneous tasks and varying label spaces.

Back to top