We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Gradient Starvation: A Learning Proclivity in Neural Networks

Abstract: We identify and formalize a fundamental gradient descent phenomenon resulting in a learning proclivity in over-parameterized neural networks. Gradient Starvation arises when cross-entropy loss is minimized by capturing only a subset of features relevant for the task, despite the presence of other predictive features that fail to be discovered. This work provides a theoretical explanation for the emergence of such feature imbalance in neural networks. Using tools from Dynamical Systems theory, we identify simple properties of learning dynamics during gradient descent that lead to this imbalance, and prove that such a situation can be expected given certain statistical structure in training data. Based on our proposed formalism, we develop guarantees for a novel regularization method aimed at decoupling feature learning dynamics, improving accuracy and robustness in cases hindered by gradient starvation. We illustrate our findings with simple and real-world out-of-distribution (OOD) generalization experiments.
Comments: Proceeding of NeurIPS 2021
Subjects: Machine Learning (cs.LG); Dynamical Systems (math.DS); Machine Learning (stat.ML)
Cite as: arXiv:2011.09468 [cs.LG]
  (or arXiv:2011.09468v4 [cs.LG] for this version)

Submission history

From: Mohammad Pezeshki [view email]
[v1] Wed, 18 Nov 2020 18:52:08 GMT (15519kb,D)
[v2] Mon, 23 Nov 2020 18:17:22 GMT (4412kb,D)
[v3] Tue, 26 Oct 2021 01:42:31 GMT (4384kb,D)
[v4] Wed, 24 Nov 2021 18:24:37 GMT (4384kb,D)

Link back to: arXiv, form interface, contact.