Exploiting Connections between Lipschitz Structures for Certifiably Robust Deep Equilibrium Models
Authors: Aaron Havens, Alexandre Araujo, Siddharth Garg, Farshad Khorrami, Bin Hu
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We also support our theoretical understanding with new empirical results, which show that our proposed method improves the certified robust accuracy of DEQs on classification tasks. All codes and experiments are made available at https://github.com/Aaron Havens/Exploiting Lipschitz DEQ. and 4 Numerical Experiments |
| Researcher Affiliation | Academia | Aaron J. Havens 1 Alexandre Araujo 2 Siddharth Garg2 Farshad Khorrami2 Bin Hu1 1 CSL & ECE, University of Illinois Urbana-Champaign 2 ECE, New York University |
| Pseudocode | No | The paper does not contain any block explicitly labeled Pseudocode or Algorithm. |
| Open Source Code | Yes | All codes and experiments are made available at https://github.com/Aaron Havens/Exploiting Lipschitz DEQ. |
| Open Datasets | Yes | For the experiments, we trained Mon DEQ and SLL networks that would serve as initialization for LBEN. and First, we present a small-scale experiment on the MNIST dataset with Lip-Mon DEQ and LBEN initialized from the trained Lip-Mon DEQ. and We now present some experiments on CIFAR10 and CIFAR100 datasets [20]. |
| Dataset Splits | No | For the experiments, we trained Mon DEQ and SLL networks that would serve as initialization for LBEN. and First, we present a small-scale experiment on the MNIST dataset with Lip-Mon DEQ and LBEN initialized from the trained Lip-Mon DEQ. and We now present some experiments on CIFAR10 and CIFAR100 datasets [20]. The paper mentions datasets used but does not explicitly state the training, validation, or test split percentages or methodology. |
| Hardware Specification | No | The paper does not specify any particular hardware (e.g., GPU or CPU models) used for running the experiments. It only mentions computational cost as a reason for small networks, but no specifics. |
| Software Dependencies | No | We use the Re LU nonlinearity which is slope-restricted on [0, 1]. The paper mentions the activation function but does not specify any software dependencies with version numbers (e.g., PyTorch, TensorFlow, Python version). |
| Experiment Setup | Yes | To fine-tune the LBEN after initialization, we use a small learning rate of 1e 7 with a ℓ1-regularization of 0.1 during 40 epochs (for more details, see the publicly available code). and For the SLL network (e.g., (2) in Table 1), we use 4 convolutional layers with circular padding and 2 dense layers. The convolutional and dense layers have 8 and 512 channels/features, respectively. |