DECision-Aid Tools Transfer Learning OrieNted

Transfer Learning (TL) seeks to improve the learning of a new target task by leveraging prior knowledge from similar source tasks. Its growing popularity within the machine learning community and industry stems from its ability to deliver strong performance in situations where data is scarce, diversity is needed, or frugality is essential.

While numerous statistical and algorithmic methods have been developed to address what and how to transfer, their effectiveness is often evaluated through numerical experiments with limited theoretical guarantees. DECATTLON takes a novel approach by focusing on when to transfer knowledge, aiming to create mathematical tools that guide decision-making across three distinct TL scenarios:

1. Quantifying transferability: Defining a quantitative measure and a statistical test to assess whether applying transfer is appropriate in a non-parametric regression setting.

2. Adapting transfer learning to systems that evolve over time, such as in stochastic differential equations or diffusion models.