Current browse context:
math.ST
Change to browse by:
References & Citations
Mathematics > Statistics Theory
Title: Convergence rates of deep ReLU networks for multiclass classification
(Submitted on 2 Aug 2021)
Abstract: For classification problems, trained deep neural networks return probabilities of class memberships. In this work we study convergence of the learned probabilities to the true conditional class probabilities. More specifically we consider sparse deep ReLU network reconstructions minimizing cross-entropy loss in the multiclass classification setup. Interesting phenomena occur when the class membership probabilities are close to zero. Convergence rates are derived that depend on the near-zero behaviour via a margin-type condition.
Submission history
From: Johannes Schmidt-Hieber [view email][v1] Mon, 2 Aug 2021 15:14:12 GMT (131kb)
Link back to: arXiv, form interface, contact.