Asymptotic Analysis of Meta-learning as a Recommendation Problem - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year : 2021

Asymptotic Analysis of Meta-learning as a Recommendation Problem

Abstract

Meta-learning tackles various means of learning from past tasks to perform new tasks better. In this paper, we focus on one particular statement of meta-learning: learning to recommend algorithms. We focus on a finite number of algorithms, which can be executed on tasks drawn i.i.d. according to a "meta-distribution". We are interested in generalization performance of meta-predict strategies, i.e., the expected algorithm performances on new tasks drawn from the same meta-distribution. Assuming the perfect knowledge of the meta-distribution (i.e., in the limit of a very large number of training tasks), we ask ourselves under which conditions algorithm recommendation can benefit from meta-learning, and thus, in some sense, "defeat" the No-Free-Lunch theorem. We analyze four meta-predict strategies: Random, Mean, Greedy and Optimal. We identify optimality conditions for such strategies. We also define a notion of meta-learning complexity as the cardinal of the minimal clique of complementary algorithms. We illustrate our findings on experiments conducted on artificial and real data.
Fichier principal
Vignette du fichier
PMLR_AAAI2020_Meta_learning_paper___draft_version_0022.pdf (2.06 Mo) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-03098180 , version 1 (05-01-2021)
hal-03098180 , version 2 (01-12-2021)
hal-03098180 , version 3 (08-01-2022)

Identifiers

  • HAL Id : hal-03098180 , version 3

Cite

Zhengying Liu, Isabelle Guyon. Asymptotic Analysis of Meta-learning as a Recommendation Problem. AAAI Workshop on Meta-Learning and MetaDL, Feb 2021, Virtual, Canada. ⟨hal-03098180v3⟩
104 View
77 Download

Share

Gmail Mastodon Facebook X LinkedIn More