Learning How to Learn is Learning With Point Sets
Tom Minka, Rosalind W. Picard · 1999
It has been proposed that learning how to learn be understood in terms of "learning a prior," from a Bayesian point-of-view (Baxter, 1996b). This paper presents an alternative interpretation: learning how to learn is ordinary learning but on point sets, rather than points. The idea behind learning how to learn is to partition the data, learn a model for the partitions, and then apply this model to new partitions. Ordinary learning methods do the same thing but with individual data points as the partitions. The partitioning for learning how to learn may be recovered automatically and may be applied recursively, leading to "task clustering" models. Virtually all existing approaches fit naturally into this unifying framework, including learning a distance metric and learning internal representations. 1 INTRODUCTION In the classical learning scenario, we have some independent and identically distributed (IID) samples of a function and want to complete the function, i.e. guess the values ...