A Dynamical System Perspective for Lipschitz Neural Networks
Authors: Laurent Meunier, Blaise J Delattre, Alexandre Araujo, Alexandre Allauzen
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | A comprehensive set of experiments on several datasets demonstrates the scalability of our architecture and the benefits as an ℓ2-provable defense against adversarial examples. |
| Researcher Affiliation | Collaboration | 1Miles Team, LAMSADE, Université Paris-Dauphine, PSL University, Paris, France 2Meta AI Research, Paris, France 3Foxstream, Lyon, France 4INRIA, Ecole Normale Supérieure, CNRS, PSL University, Paris, France 5ESPCI, Paris, France. |
| Pseudocode | Yes | Algorithm 1 Computation of a Convex Potential Layer |
| Open Source Code | Yes | Our code is available at https://github.com/ MILES-PSL/Convex-Potential-Layer |
| Open Datasets | Yes | We demonstrate the effectiveness of our approach on a classification task with CIFAR10 and CIFAR100 datasets (Krizhevsky et al., 2009). |
| Dataset Splits | No | The paper states the use of CIFAR10 and CIFAR100 datasets for experiments, which have standard training and test splits, but it does not explicitly describe the creation or size of a validation split (e.g., percentages or sample counts) used for model selection or hyperparameter tuning. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments, such as GPU models, CPU types, or cloud computing specifications. |
| Software Dependencies | No | The paper mentions leveraging 'Py Torch framework (Paszke et al., 2019)' but does not specify its version number or any other software dependencies with their versions. |
| Experiment Setup | Yes | We trained our networks with a batch size of 256 over 200 epochs. We use standard data augmentation (i.e., random cropping and flipping), a learning rate of 0.001 with Adam optimizer (Diederik P. Kingma, 2014) without weight decay and a piecewise triangular learning rate scheduler. We used a margin loss5 with margin parameter set to 0.7. |