Robustness Implies Generalization via Data-Dependent Generalization Bounds

Authors: Kenji Kawaguchi, Zhun Deng, Kyle Luh, Jiaoyang Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental The experiments on real-world data and theoretical models demonstrate near-exponential improvements in various situations. To achieve these improvements, we do not require additional assumptions on the unknown distribution; instead, we only incorporate an observable and computable property of the training samples. A key technical innovation is an improved concentration bound for multinomial random variables that is of independent interest beyond robustness and generalization.
Researcher Affiliation Academia 1National University of Singapore 2Harvard University 3University of Colorado Boulder 4New York University.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes For real-world data, we adopted the standard benchmark datasets: MNIST (Le Cun et al., 1998), CIFAR-10 and CIFAR-100 (Krizhevsky & Hinton, 2009), SVHN (Netzer et al., 2011), Fashion-MNIST (FMNIST) (Xiao et al., 2017), Kuzushiji-MNIST (KMNIST) (Clanuwat et al., 2019), and Semeion (Srl & Brescia, 1994).
Dataset Splits Yes Following the literature on semi-supervised learning, we split the training data points into labeled data points (500 for Semeion and 5000 for all other datasets) and unlabeled data points (the remainder of the training data).
Hardware Specification No The paper does not provide specific hardware details used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers.
Experiment Setup Yes The data space is normalized such that X [0, 1]d for the dimensionality d of each input data. Accordingly, we used the infinity norm and a diameter of 0.1 for the ϵ-covering in all experiments.