Can Implicit Bias Imply Adversarial Robustness?

Authors: Hancheng Min, Rene Vidal

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Moreover, numerical experiments on real datasets show that shallow networks with our generalized Re LU activation functions are much more robust than those with a Re LU activation.
Researcher Affiliation Academia 1University of Pennsylvania, Philadelphia, PA, USA. Correspondence to: Hancheng Min <hanchmin@seas.upenn.edu>.
Pseudocode Yes Algorithm 1 Estimating dist(fp, F)
Open Source Code Yes Code available at https://github.com/hanchmin/pReLU_ICML24.
Open Datasets Yes We first consider training a p Re LU network of width h = 500 to predict whether an MNIST digit is even or odd.
Dataset Splits No We run SGD (batch size 100 and step size 0.2 for 2 105 epochs) with small initialization (all weights initialized as mean-zero Gaussian with standard deviation 10 7) to train a p Re LU network with h = 2000 neurons on a dataset drawn from DX,Y with D = 1000, K = 10, and K1 = 6.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory, or cloud services) are mentioned in the paper.
Software Dependencies No The paper mentions optimizers like "SGD" and "Adam," but does not provide specific version numbers for any software, programming languages, or libraries used for implementation.
Experiment Setup Yes We run SGD (batch size 100 and step size 0.2 for 2 105 epochs) with small initialization (all weights initialized as mean-zero Gaussian with standard deviation 10 7) to train a p Re LU network with h = 2000 neurons on a dataset drawn from DX,Y with D = 1000, K = 10, and K1 = 6.