PAC-Bayesian Bounds for Randomized Empirical Risk Minimizers

📝 Abstract

The aim of this paper is to generalize the PAC-Bayesian theorems proved by Catoni in the classification setting to more general problems of statistical inference. We show how to control the deviations of the risk of randomized estimators. A particular attention is paid to randomized estimators drawn in a small neighborhood of classical estimators, whose study leads to control the risk of the latter. These results allow to bound the risk of very general estimation procedures, as well as to perform model selection.

💡 Analysis

The aim of this paper is to generalize the PAC-Bayesian theorems proved by Catoni in the classification setting to more general problems of statistical inference. We show how to control the deviations of the risk of randomized estimators. A particular attention is paid to randomized estimators drawn in a small neighborhood of classical estimators, whose study leads to control the risk of the latter. These results allow to bound the risk of very general estimation procedures, as well as to perform model selection.

📄 Content

Ò Ø ÓÒ ½º½º Ä Ø Ù× ÔÙØ

Ü ÑÔÐ ½º¿ ´ Ò× ØÝ ×Ø Ñ Ø ÓÒµº À Ö ¸Û ××ÙÑ Ø Ø P 1 = … = P N = P Ò ÓÒ× ÕÙ ÒØÐÝ Ø Ø

.

P exp sup

´ º½µ

´ º¾µ

´ º¿µ

´ º µ β -

, θi,β -βr ′ (., θ)

-log π i exp -βr ′ (., θ) .

´ º½¿µ

´ º½ µ R ′ θt , θt j ≤ B(t j , t s(j) ) + B(t s(j) , t j ) + B( t, t ŝ) + B(t ŝ, t) .

Ï Ö Ó Ò ØÓ ÙÔÔ Ö ÓÙÒ × Ô Ö Ø ÐÝ B(t j , t s(j) ) + B(t s(j) , t j ) Ò B( t, t ŝ) + B(t ŝ, t)º Ä Ø Ù× Ö×Ø Ð Û Ø Ø Ø ÖÑ B(t j , t s(j) ) + B(t s(j) , t j )

´ º½ µ B(t j , t s(j) ) + B(t s(j) , t j) , θt j ) ≤ B(t j , t s(j) ) + B(t s(j) , t j ) ≤ Eδ N (i, q, ε, κ).

ÈÐÙ Ò ÐÐ Ø × Ö ×ÙÐØ× ÒØÓ ÁÒ ÕÙ Ð ØÝ ´ º½ µ¸Û Ó Ø Ò¸ 1 -2λx N R ′ θt , θt j ≤ Eδ N (i, q, ε, κ) + 2λ N xEδ N (i, q, ε, κ) + 2xR ′ ( θt j , θ i ) + 2xR ′ (θ i , θ) + ϕ(x)

.

× Ù×Ù Ð¸Ð Ø Ù× ÔÔÐÝ Ä ÑÑ º ØÓ ÙÔÔ Ö ÓÙÒ C(t j )¸Ä ÑÑ º¿ ØÓ ÙÔÔ Ö ÓÙÒ R ′ ( θt j , θ i ) Ò Ä ÑÑ º½ ØÓ ÙÔÔ Ö ÓÙÒ ϕ(x)º Ä Ø Ù× ÔÙØ

θt , θt j ≤ E ′ δ N (i, q, ε, κ). Ì Ò¸ ÓÖ ÒÝ (θ, θ ′ ) ∈ Θ 2 Ŗ(θ)

This content is AI-processed based on ArXiv data.