Gradient methods never overfit on separable data

MaRDI QIDQ4999011zbMATH OpenFDO

Authors Ohad Shamir

Publication date 9 July 2021

Full work available at URL https://arxiv.org/abs/2007.00028, https://jmlr.csail.mit.edu/papers/v22/20-997.html

zbMATH Keywords

gradient methodslinear predictorsimplicit biaslinearly separable datamargin-based generalization bounds

Mathematics Subject Classification ID

Learning and adaptive systems in artificial intelligence (68T05)

Recommendations

• The implicit bias of gradient descent on separable data
• Bias of homotopic gradient descent for the hinge loss
• scientific article; zbMATH DE number 7370646
• Tractability from overparametrization: the example of the negative perceptron

Cites work

• Lectures on convex optimization
• Optimality of SVM: novel proofs and tighter bounds
• Simplified PAC-Bayesian margin bounds.
• The implicit bias of gradient descent on separable data
• Understanding machine learning. From theory to algorithms

Cited in

(1)

• Tractability from overparametrization: the example of the negative perceptron

This page was built for publication: Gradient methods never overfit on separable data

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4999011)