Generalization Bounds with Logarithmic Negative-Sample Dependence for Adversarial Contrastive Learning

Authors: Naghmeh Ghanooni, Waleed Mustafa, Yunwen Lei, Anthony Widjaja Lin, Marius Kloft

TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our theoretical results are supported by experiments on real-world datasets. Empirical evaluations are provided in Section 6. 6 Experiments We evaluate our theoretical results to two widely used benchmark datasets from the image domain: CIFAR-10 and CIFAR-100 (Krizhevsky, 2009).
Researcher Affiliation Academia Naghmeh Ghanooni EMAIL Department of Computer Science RPTU, Kaiserslautern, Germany Waleed Mustafa EMAIL Department of Computer Science RPTU, Kaiserslautern, Germany Yunwen Lei EMAIL Department of Mathematics University of Hong Kong, Hong Kong, China Anthony Widjaja Lin EMAIL MPI-SWS RPTU, Kaiserslautern, Germany Marius Kloft EMAIL Department of Computer Science RPTU, Kaiserslautern, Germany
Pseudocode No The paper describes theoretical framework and experimental results but does not contain any clearly labeled pseudocode or algorithm blocks. Methods are described textually and mathematically.
Open Source Code No The paper does not contain any explicit statements or links regarding the release of source code for the methodology described.
Open Datasets Yes We evaluate our theoretical results to two widely used benchmark datasets from the image domain: CIFAR-10 and CIFAR-100 (Krizhevsky, 2009). CIFAR-10 and CIFAR-100 consist of 50,000 training images and 10,000 testing images, organized into 10 and 100 classes, respectively.
Dataset Splits Yes CIFAR-10 and CIFAR-100 consist of 50,000 training images and 10,000 testing images, organized into 10 and 100 classes, respectively.
Hardware Specification No The paper mentions that models are trained but does not provide any specific details about the hardware used (e.g., GPU models, CPU types, or memory specifications).
Software Dependencies No Both models are trained using the Adam optimizer, with a learning rate of 1e-3. To generate adversarial perturbations, we employ the ℓ PGD algorithm (Madry, 2017) with a step size of ϵ/255. The paper mentions software components like 'Adam optimizer' and 'PGD algorithm' but does not provide specific version numbers for any software libraries or frameworks (e.g., PyTorch, TensorFlow, Python version).
Experiment Setup Yes Both models are trained using the Adam optimizer, with a learning rate of 1e-3. To generate adversarial perturbations, we employ the ℓ PGD algorithm (Madry, 2017) with a step size of ϵ/255, where ϵ represents the maximum allowable perturbation. We calculate the generalization error for different values of ϵ = {2/255, 4/255, 8/255, 16/255, 32/255, 64/255, 128/255}, with varying negative samples, K = {63, 127, 511, 1023}.