Learning Curves for SGD on Structured Features

Authors: Blake Bordelon, Cengiz Pehlevan

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the accuracy of our theory on random feature models and wide neural networks trained with SGD on real datasets such as MNIST and CIFAR-10.
Researcher Affiliation Academia Blake Bordelon & Cengiz Pehlevan John A. Paulson School of Engineering and Applied Sciences Center for Brain Science Harvard University Cambridge, MA 02138, USA {blake bordelon,cpehlevan}@g.harvard.edu
Pseudocode No No pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes REPRODUCIBILITY STATEMENT The code to reproduce the experimental components of this paper can be found here https://github.com/Pehlevan-Group/sgd_structured_features, which contains jupyter notebook files which we ran in Google Colab.
Open Datasets Yes We demonstrate the accuracy of our theory on random feature models and wide neural networks trained with SGD on real datasets such as MNIST and CIFAR-10.
Dataset Splits No The paper mentions using 'training points' and 'test set' but does not specify validation splits or other detailed splitting methodology.
Hardware Specification No The paper mentions that experiments were run in 'Google Colab' but does not provide specific hardware details (e.g., GPU/CPU models, memory).
Software Dependencies No The paper mentions using 'Neural Tangents API' but does not provide specific version numbers for software dependencies.
Experiment Setup Yes We explore in detail the effect of minibatch size, m, on learning dynamics. By varying m, we can interpolate our theory between single sample SGD (m = 1) and gradient descent on the population loss (m ).