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

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

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

Computer Science > Machine Learning

Title: SGD with Hardness Weighted Sampling for Distributionally Robust Deep Learning

Abstract: Distributionally Robust Optimization (DRO) has been proposed as an alternative to Empirical Risk Minimization (ERM) in order to account for potential biases in the training data distribution. However, its use in deep learning has been severely restricted due to the relative inefficiency of the optimizers available for DRO in comparison to the wide-spread Stochastic Gradient Descent (SGD) based optimizers for deep learning with ERM. We propose SGD with Hardness weighted sampling, an efficient optimization method for machine learning with DRO with a focus on deep learning. In this work, we propose SGD with hardness weighted sampling, a principled and efficient optimization method for DRO in machine learning that is particularly suited in the context of deep learning. We show that our optimization method can be interpreted as a principled Hard Example Mining strategy. Similar to an online hard example mining strategy in essence and in practice, the proposed algorithm is straightforward to implement and computationally as efficient as SGD-based optimizers used for deep learning. It only requires adding a softmax layer and maintaining a history of the loss values for each training example to compute adaptive sampling probabilities. In contrast to typical ad hoc hard mining approaches, and exploiting recent theoretical results in deep learning optimization, we We also prove the convergence of our DRO algorithm for over-parameterized deep learning networks with ReLU activation and finite number of layers and parameters. Preliminary results demonstrate the feasibility and usefulness of our approach.
Subjects: Machine Learning (cs.LG); Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2001.02658 [cs.LG]
  (or arXiv:2001.02658v1 [cs.LG] for this version)

Submission history

From: Lucas Fidon [view email]
[v1] Wed, 8 Jan 2020 18:02:56 GMT (138kb,D)
[v2] Thu, 25 Jun 2020 09:43:22 GMT (451kb,D)
[v3] Wed, 9 Feb 2022 14:53:52 GMT (2673kb,D)
[v4] Thu, 14 Jul 2022 22:03:25 GMT (2688kb,D)

Link back to: arXiv, form interface, contact.