Understanding Gradient Descent on the Edge of Stability in Deep Learning

Authors: Sanjeev Arora, Zhiyuan Li, Abhishek Panigrahi

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

Reproducibility Variable Result LLM Response
Research Type Experimental The above theoretical results have been corroborated by an experimental study.
Researcher Affiliation Academia 1 Princeton University
Pseudocode Yes Algorithm 1 Perturbed Normalized Gradient Descent; Algorithm 3 Perturbed Gradient Descent on L
Open Source Code No The paper does not contain an explicit statement about releasing source code or a link to a code repository for the described methodology.
Open Datasets Yes We perform our experiments on a VGG-16 model (Simonyan & Zisserman, 2014) trained on CIFAR-10 dataset (Krizhevsky et al.) with Normalized GD and GD with L.
Dataset Splits No The paper mentions using a sample of training data but does not specify train/validation/test splits, percentages, or counts for any dataset.
Hardware Specification No The paper mentions training on a "single GPU" but does not specify any particular GPU model, CPU, or other hardware details.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup Yes The network had 784 hidden units, with Ge LU activation function (Hendrycks & Gimpel, 2016). We used the loss function L as the mean squared loss to ensure the existence of minimizers and thus the manifold. For efficient training on a single GPU, we consider a sample of 1000 randomly selected points from the training data.