Adversarial Robustness is at Odds with Lazy Training

Authors: Yunjuan Wang, Enayat Ullah, Poorya Mianjy, Raman Arora

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

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5 we provide empirical support for our theory, and conclude with a discussion in Section 6. 5 Experiments The primary goal of this section is to provide empirical support for our theoretical findings.
Researcher Affiliation Academia Yunjuan Wang Department of Computer Science Johns Hopkins University Baltimore, MD, 21218 ywang509@jhu.edu
Pseudocode Yes To ensure that the updates remain in the lazy regime, after each gradient step, we project the weights onto B2, (W0, V ); see Algo. 1 in the appendix for details.
Open Source Code Yes Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See supplemental material.
Open Datasets Yes We utilize the MNIST dataset for our empirical study. MNIST is a dataset of 28 28 greyscale handwritten digits [Le Cun et al., 1998].
Dataset Splits Yes We extract examples corresponding to images of the digits 0 and 1 , resulting in 12665 training examples and 2115 test examples.
Hardware Specification No The paper does not specify any particular hardware, such as GPU or CPU models, or cloud computing resources used for the experiments. The checklist explicitly states 'N/A' for 'total amount of compute and the type of resources used'.
Software Dependencies No The paper describes the algorithms used (standard SGD, PGD) but does not specify any software libraries or dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, etc.).
Experiment Setup Yes We train the network using standard SGD on the logistic loss using a learning rate of 0.1. ... We experiment with C0 {10, 20, 30, 40}. ... For this set of experiments, we also normalize the data, so that x = 1 ... We experiment with values of V and the size of δ in the intervals 0.05 V 1 and 0.05 δ 1, respectively.