[2111.04746] Realizable Studying is All You Want

0
29


Obtain a PDF of the paper titled Realizable Studying is All You Want, by Max Hopkins and three different authors

Obtain PDF
HTML (experimental)

Summary:The equivalence of realizable and agnostic learnability is a basic phenomenon in studying idea. With variants starting from classical settings like PAC studying and regression to current developments akin to adversarially strong studying, it is shocking that we nonetheless lack a unified idea; conventional proofs of the equivalence are usually disparate, and depend on robust model-specific assumptions like uniform convergence and pattern compression.

On this work, we give the primary model-independent framework explaining the equivalence of realizable and agnostic learnability: a three-line blackbox discount that simplifies, unifies, and extends our understanding throughout all kinds of settings. This consists of fashions with no identified characterization of learnability akin to studying with arbitrary distributional assumptions and extra normal loss features, in addition to a bunch of different fashionable settings akin to strong studying, partial studying, truthful studying, and the statistical question mannequin.

Extra typically, we argue that the equivalence of realizable and agnostic studying is definitely a particular case of a broader phenomenon we name property generalization: any fascinating property of a studying algorithm (e.g. noise tolerance, privateness, stability) that may be glad over finite speculation courses extends (presumably in some variation) to any learnable speculation class.

Submission historical past

From: Max Hopkins [view email]
[v1]
Mon, 8 Nov 2021 19:00:00 UTC (55 KB)
[v2]
Solar, 25 Sep 2022 08:34:25 UTC (71 KB)
[v3]
Fri, 3 Feb 2023 12:06:15 UTC (72 KB)
[v4]
Sat, 3 Feb 2024 00:55:16 UTC (98 KB)



Supply hyperlink

LEAVE A REPLY

Please enter your comment!
Please enter your name here