Scaling the Convex Barrier with Active Sets

Authors: Alessandro De Palma, Harkirat Behl, Rudy R Bunel, Philip Torr, M. Pawan Kumar

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically demonstrate the effectiveness of our method under two settings. On incomplete verification ( 5.1), we assess the speed and quality of bounds compared to other bounding algorithms. On complete verification ( 5.2), we examine whether our speed-accuracy trade-offs correspond to faster exact verification.
Researcher Affiliation Academia Alessandro De Palma , Harkirat Singh Behl , Rudy Bunel, Philip H.S. Torr, M. Pawan Kumar University of Oxford {adepalma,harkirat,phst,pawan}@robots.ox.ac.uk bunel.rudy@gmail.com
Pseudocode Yes Pseudo-code can be found in appendix D.
Open Source Code Yes Our implementation is based on Pytorch (Paszke et al., 2017) and is available at https://github.com/oval-group/scaling-the-convex-barrier.
Open Datasets Yes We next evaluate the performance on complete verification, verifying the adversarial robustness of a network to perturbations in ℓ norm on a subset of the dataset by Lu & Kumar (2020)...
Dataset Splits No The paper mentions using 'CIFAR-10 test set' and 'MNIST test set' for evaluation, and also discusses pre-trained networks. However, it does not explicitly provide details on how the training and validation sets were split for model development or fine-tuning, or if standard pre-defined splits for all three subsets were utilized consistently for the models they trained or adapted.
Hardware Specification Yes All the experiments and bounding computations (including intermediate bounds) were run on a single Nvidia Titan Xp GPU, except Gurobi-based methods and Active Set CPU. These were run on i7-6850K CPUs, utilising 4 cores for the incomplete verification experiments, and 6 cores for the more demanding complete verification experiments.
Software Dependencies No The paper mentions 'Pytorch' and 'Gurobi' but does not specify their version numbers or other software dependencies with version numbers.
Experiment Setup Yes For Big-M, replicating the findings by Bunel et al. (2020a) on their supergradient method, we linearly decrease the step size from 10 2 to 10 4. Active Set is initialized with 500 Big-M iterations, after which the step size is reset and linearly scaled from 10 3 to 10 6. We found the addition of variables to the active set to be effective before convergence: we add variables every 450 iterations, without re-scaling the step size again. Every addition consists of 2 new variables...