BNN-DP: Robustness Certification of Bayesian Neural Networks via Dynamic Programming

Authors: Steven Adams, Andrea Patane, Morteza Lahijanian, Luca Laurenti

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On a set of experiments on various regression and classification tasks and BNN architectures, we show that BNN-DP outperforms state-of-the-art methods by up to four orders of magnitude in both tightness of the bounds and computational efficiency.We validate our framework on several regression and classification tasks, including the Kin8nm, MNIST, Fashion MNIST, and CIFAR-10 datasets, and a range of BNN architectures.
Researcher Affiliation Academia Delft Center for Systems and Control, Technical University of Delft, Delft, 2628 CD, The Netherlands School of Computer Science and Statistics, Trinity College Dublin, Dublin 2, Ireland Departement of Aerospace Engineering Sciences and Computer Science, University of Colorado Boulder, Boulder, CO 80303, USA.
Pseudocode Yes Algorithm 1 Adversarial Robustness for Classification; Algorithm 2 Interval relaxation procedure for Eqn. 5a with h = softmax; Algorithm 3 Back-Propagation of PWA Relaxations.; Algorithm 4 Adversarial Robustness for Regression
Open Source Code Yes Our code is available at https://github.com/ sjladams/BNN_DP.
Open Datasets Yes We validate our framework on several regression and classification tasks, including the Kin8nm, MNIST, Fashion MNIST, and CIFAR-10 datasets, and a range of BNN architectures. The latter dataset contains state-space readings for the dynamics of an 8 link robot arm, and is commonly used as a regression task to benchmark BNNs (Hern andez-Lobato & Adams, 2015; Gal & Ghahramani, 2016).10Available at http://www.cs.toronto.edu/ delve.
Dataset Splits No The paper refers to testing on '100 test points' and '100 randomly sampled test points' but does not explicitly provide training, validation, and test dataset splits with percentages or sample counts.
Hardware Specification Yes We ran our experiments on an AMD EPYC 7252 8-core CPU and train the BNNs using Noisy Adam (Zhang et al., 2018a) and variational online Gauss-Newton (Khan et al., 2018).
Software Dependencies No The paper mentions using 'Noisy Adam' and 'variational online Gauss-Newton' as training methods but does not list specific software packages with version numbers (e.g., PyTorch 1.9, TensorFlow 2.x).
Experiment Setup No The paper mentions training methods (Noisy Adam, variational online Gauss-Newton) and BNN architectures, but does not provide specific hyperparameter values such as learning rates, batch sizes, or number of epochs.