Can overfitted deep neural networks in adversarial training generalize? – An approximation viewpoint

In this talk, I will discuss whether overfitted DNNs in adversarial training can generalize from an approximation viewpoint. We prove by construction the existence of infinitely many adversarial training classifiers on over-parameterized DNNs that obtain arbitrarily small adversarial training error (overfitting), whereas achieving good robust generalization error under certain conditions concerning the data quality, well separated, and perturbation level. This construction is optimal and thus points out the fundamental limits of DNNs under adversarial training with statistical guarantees. Part of this talk comes from our recent work.

Format
Movie
Episodes
1
Episode Duration
52 mins
Status
Finished
All details and tags

Details

Type
Movie
Format
Movie
Episodes
1
Episode Duration
52 mins
Status
Finished
Start Date
Mar 1, 2024
Season
2024
Ratings
0
Popularity
0
Favorites
0
Reviews
0
Producers
University of Warwick
Genres
Documentary, Crime
Country
GB

Description

In this talk, I will discuss whether overfitted DNNs in adversarial training can generalize from an approximation viewpoint. We prove by construction the existence of infinitely many adversarial training classifiers on over-parameterized DNNs that obtain arbitrarily small adversarial training error (overfitting), whereas achieving good robust generalization error under certain conditions concerning the data quality, well separated, and perturbation level. This construction is optimal and thus points out the fundamental limits of DNNs under adversarial training with statistical guarantees. Part of this talk comes from our recent work.

Immersion

Japanese

Characters

Staff

Recommendations

No recommendations yet. Add the first one.