Understanding Certified Training with Interval Bound Propagation
Authors: Yuhao Mao, Mark Niklas Mueller, Marc Fischer, Martin Vechev
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we show how these results on DLNs transfer to ReLU networks, before conducting an extensive empirical study, (i) confirming this transferability and yielding state-of-the-art certified accuracy..." and Section 4 "EMPIRICAL EVALUATION ANALYSIS". |
| Researcher Affiliation | Academia | Department of Computer Science, ETH Zürich, Swizterland {yuhao.mao, mark.mueller, marc.fischer, martin.vechev}@inf.ethz.ch |
| Pseudocode | No | No pseudocode or clearly labeled algorithm block was found in the paper. |
| Open Source Code | Yes | We publish our code, trained models, and detailed instructions on how to reproduce our results at https://github.com/eth-sri/ibp-propagation-tightness. |
| Open Datasets | Yes | We use the MNIST (Le Cun et al., 2010) and CIFAR-10 (Krizhevsky et al., 2009) datasets for our experiments. Both are open-source and freely available. |
| Dataset Splits | Yes | We use the MNIST (Le Cun et al., 2010) and CIFAR-10 (Krizhevsky et al., 2009) datasets for our experiments. Both are open-source and freely available. |
| Hardware Specification | No | The paper does not provide specific details regarding the hardware (e.g., CPU, GPU models, or memory specifications) used for running the experiments. |
| Software Dependencies | No | The paper mentions providing code but does not specify software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions) in the text. |
| Experiment Setup | Yes | Specifically, for MNIST, the first 20 epochs are used for ϵ-scheduling, increasing ϵ smoothly from 0 to the target value. Then, we train an additional 50 epochs with two learning rate decays of 0.2 at epochs 50 and 60, respectively. For CIFAR-10, we use 80 epochs for ϵ-annealing, after training models with standard training for 1 epoch. We continue training for 80 further epochs with two learning rate decays of 0.2 at epochs 120 and 140, respectively. The initial learning rate is 5e-3 and the gradients are clipped to an L2 norm of at most 10.0 before every step. |