In a multitude of potential AI applications, from precision medicine through to automated fraud detection, we are faced with a problem of data scarcity, whereby the practitioner only has access to a small quantity of data closely aligned to their task of interest. Classical approaches to Machine Learning are ill-suited to such situations since they either require a very large data-set with the same underlying population as their task of interest, or require a tightly constrained model class which damages their flexibility. Humans, on the other hand, learn flexible patterns from scarce data continuously, by leveraging symmetries with distinct but related tasks of interest. Transfer learning mirrors this process by combining information from a relatively small task-specific data set from a target-distribution, with a larger data set from a distinct but related source-distribution. An important challenge in this domain is to identify techniques capable of automatically adapting to the unknown characteristics of the problem such as the strength of the relationship between the source and target distributions. In our recent joint paper with Timothy Cannings, at the University of Edinburgh, and Richard Samworth, at the University of Cambridge, we introduce an algorithm for Adaptive Transfer Learning which provably adapts to these unknown characteristics in a provably optimal way.
Please follow these links to find out more:
Institute of Mathematical Statistics/Annals of Statistics Future Papers