Abstract
We derive upper bounds on the generalization error of a learning algorithm in terms of the mutual information between its input and output. The bounds provide an information-theoretic understanding of generalization in learning problems, and give theoretical guidelines for striking the right balance between data fit and generalization by controlling the input-output mutual information. We propose a number of methods for this purpose, among which are algorithms that regularize the ERM algorithm with relative entropy or with random noise. Our work extends and leads to nontrivial improvements on the recent results of Russo and Zou.
| Original language | English (US) |
|---|---|
| Pages (from-to) | 2525-2534 |
| Number of pages | 10 |
| Journal | Advances in Neural Information Processing Systems |
| Volume | 2017-December |
| State | Published - Jan 1 2017 |
| Event | 31st Annual Conference on Neural Information Processing Systems, NIPS 2017 - Long Beach, United States Duration: Dec 4 2017 → Dec 9 2017 |
Fingerprint
ASJC Scopus subject areas
- Computer Networks and Communications
- Information Systems
- Signal Processing
Cite this
Information-theoretic analysis of generalization capability of learning algorithms. / Xu, Aolin; Raginsky, Maxim.
In: Advances in Neural Information Processing Systems, Vol. 2017-December, 01.01.2017, p. 2525-2534.Research output: Contribution to journal › Conference article
}
TY - JOUR
T1 - Information-theoretic analysis of generalization capability of learning algorithms
AU - Xu, Aolin
AU - Raginsky, Maxim
PY - 2017/1/1
Y1 - 2017/1/1
N2 - We derive upper bounds on the generalization error of a learning algorithm in terms of the mutual information between its input and output. The bounds provide an information-theoretic understanding of generalization in learning problems, and give theoretical guidelines for striking the right balance between data fit and generalization by controlling the input-output mutual information. We propose a number of methods for this purpose, among which are algorithms that regularize the ERM algorithm with relative entropy or with random noise. Our work extends and leads to nontrivial improvements on the recent results of Russo and Zou.
AB - We derive upper bounds on the generalization error of a learning algorithm in terms of the mutual information between its input and output. The bounds provide an information-theoretic understanding of generalization in learning problems, and give theoretical guidelines for striking the right balance between data fit and generalization by controlling the input-output mutual information. We propose a number of methods for this purpose, among which are algorithms that regularize the ERM algorithm with relative entropy or with random noise. Our work extends and leads to nontrivial improvements on the recent results of Russo and Zou.
UR - http://www.scopus.com/inward/record.url?scp=85047009011&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85047009011&partnerID=8YFLogxK
M3 - Conference article
AN - SCOPUS:85047009011
VL - 2017-December
SP - 2525
EP - 2534
JO - Advances in Neural Information Processing Systems
JF - Advances in Neural Information Processing Systems
SN - 1049-5258
ER -