Set-Based Training for Neural Network Verification
Authors: Lukas Koller, Tobias Ladner, Matthias Althoff
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive evaluation demonstrates that set-based training produces robust neural networks with competitive performance, which can be verified using fast (polynomial-time) verification algorithms due to the reduced output set. An extensive empirical evaluation in which we demonstrate the competitive performance of our set-based training and compare it with state-of-the-art robust training approaches. Moreover, we include large-scale ablation studies to justify our design choices. |
| Researcher Affiliation | Academia | Lukas Koller EMAIL Technical University of Munich; Tobias Ladner EMAIL Technical University of Munich; Matthias Althoff EMAIL Professorship for Cyber-Physical Systems Technical University of Munich |
| Pseudocode | Yes | Algorithm 1: Image Enclosure of a Nonlinear Layer. Algorithm 2: Set-based Training Iteration. |
| Open Source Code | No | We use the MATLAB toolbox CORA (Althoff, 2015) to implement set-based training. This statement indicates the use of a third-party toolbox, not the release of the authors' own implementation code for their methodology. No explicit statement of code release or repository link is found. |
| Open Datasets | Yes | We train a 6-layer convolutional neural network on Mnist (Le Cun et al., 2010), Svhn (Netzer et al., 2011), Cifar-10 (Krizhevsky, 2009), and Tiny Image Net (Le & Yang, 2015). |
| Dataset Splits | Yes | We use the canonical split of training and test data for each dataset and the entire test data for evaluation; because test labels are not available for Tiny Image Net, we follow (Müller et al., 2023) and use the validation set for testing. |
| Hardware Specification | Yes | Our experiments were run on a server with 2 AMD EPYC 7763 (64 cores/128 threads), 2TB RAM, and a NVIDIA A100 40GB GPU. |
| Software Dependencies | No | The paper mentions using the 'MATLAB toolbox CORA (Althoff, 2015)' and 'Adam optimizer (Kingma & Ba, 2015)' but does not specify version numbers for these or any other software components. |
| Experiment Setup | Yes | Table 6: Training hyperparameters. #Epochs Dataset η ϵ τ Batch Size (warm-up / ramp-up) Decay ...; The weights and biases are initialized as in (Shi et al., 2021).; We use Adam optimizer (Kingma & Ba, 2015) with the recommended hyperparameters.; For any PGD during training (...) we used the settings from (Müller et al., 2023): 8 iterations with an initial step size 0.5, which is decayed twice by 0.1 at iterations 4 and 7. All PGD attacks for testing are computed with 40 iterations of step size 0.01. |