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

Download:

Current browse context:

stat.ML

Change to browse by:

References & Citations

Bookmark

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

Statistics > Machine Learning

Title: Learning Curves for SGD on Structured Features

Abstract: The generalization performance of a machine learning algorithm such as a neural network depends in a non-trivial way on the structure of the data distribution. To analyze the influence of data structure on test loss dynamics, we study an exactly solveable model of stochastic gradient descent (SGD) on mean square loss which predicts test loss when training on features with arbitrary covariance structure. We solve the theory exactly for both Gaussian features and arbitrary features and we show that the simpler Gaussian model accurately predicts test loss of nonlinear random-feature models and deep neural networks trained with SGD on real datasets such as MNIST and CIFAR-10. We show that the optimal batch size at a fixed compute budget is typically small and depends on the feature correlation structure, demonstrating the computational benefits of SGD with small batch sizes. Lastly, we extend our theory to the more usual setting of stochastic gradient descent on a fixed subsampled training set, showing that both training and test error can be accurately predicted in our framework on real data.
Comments: Added new analysis of optimal batchsize and learning rate. Provided theoretical learning curves for case where test and training measures are different and apply to predicting errors for test/train splits on real datasets. Also provided a new bound for non-Gaussian features based on a regularity condition proposed by Varre et al 2021 arXiv:2102.03183
Subjects: Machine Learning (stat.ML); Machine Learning (cs.LG)
Cite as: arXiv:2106.02713 [stat.ML]
  (or arXiv:2106.02713v3 [stat.ML] for this version)

Submission history

From: Blake Bordelon [view email]
[v1] Fri, 4 Jun 2021 20:48:20 GMT (1827kb,D)
[v2] Wed, 9 Jun 2021 14:44:31 GMT (1827kb,D)
[v3] Fri, 15 Oct 2021 12:57:28 GMT (2039kb,D)

Link back to: arXiv, form interface, contact.