A ROBUST DIFFERENTIAL NEURAL ODE OPTIMIZER
Authors: Panagiotis Theodoropoulos, Guan-Horng Liu, Tianrong Chen, Augustinos D Saravanos, Evangelos Theodorou
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, the robustness of the proposed optimizer is demonstrated through greater robust accuracy compared to benchmark optimizers when trained on clean images. Additionally, its ability to provide a performance increase when adapted to an already existing adversarial defense technique is also illustrated. |
| Researcher Affiliation | Academia | Georgia Institute of Technology, USA {ptheodor3, ghliu, tianrong.chen, asaravanos3, evangelos.theodorou}@gatech.edu |
| Pseudocode | Yes | Algorithm 1 GTSONO |
| Open Source Code | No | The paper does not contain any explicit statements about making the source code available or provide a link to a code repository. |
| Open Datasets | Yes | Datasets We carry out our experiments on two image datasets: CIFAR and SVHN. Both datasets have been standardized and consist of 3 32 32 colour images, and 10 label classes. |
| Dataset Splits | No | CIFAR10 contains 50000 training images and 10000 test images, whereas SVHN contains 73257 digits for training and 26032 for testing. The paper specifies the number of training and test images but does not explicitly mention a validation split or how one was handled. |
| Hardware Specification | Yes | All experiments are conducted on a TITAN RTX. |
| Software Dependencies | No | The ODE solver we used for every experiment is the standard Runge Kutta 4(5) adaptive solver (dopri5; Dormand & Prince (1980)) implemented by the torchdiffeq package, with the numerical tolerance set to 1e-3, and fixed integration time [0, 1] for all experiments. All experiments are conducted on the same GPU machine (TITAN RTX) and implemented with pytorch. While some software is mentioned (torchdiffeq, pytorch), specific version numbers are not provided. |
| Experiment Setup | Yes | The training for all optimizers was carried out using only non-perturbed images, with a batch size of 500 images on every dataset. The networks trained on both datasets was trained for 15 epochs. |