Learning to Learn

Learning to Learn

As noted by John Langford, the  fact that new tasks most often call for developing new ML algorithms suggests that the community still misses a general understanding of what learning is; such a general vision also is at the core of the general AI approach advocated by [LeC22]. Transfer learning, multi-task learning, out-of-distribution learning, AutoML (i.e. the automatic selection of a (quasi) optimal ML pipeline for the problem at hand) all rely on a new definition of generalization, a new perception of the potential differences between the training and test distributions, and how the former one should shape the learning process. These issues have been considered in H. Rakotoarison’s PhD (AutoML; PhD award, 2021-STIC-Univ. Paris-Saclay) and S. Yang’s post-doc (out-of-distribution learning).

… more to come

Comments are closed.