Zonotope Domains for Lagrangian Neural Network Verification

Authors: Matt Jordan, Jonathan Hayase, Alex Dimakis, Sewoong Oh

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply Zono Dual to a variety of networks trained on MNIST and CIFAR-10 and demonstrate that Zono Dual outperforms the linear programming relaxation in both tightness and runtime, and yields a tighter bounding algorithm than the prior dual approaches.
Researcher Affiliation Academia Matt Jordan UT Austin mjordan@cs.utexas.edu Jonathan Hayase University of Washington jhayase@cs.washington.edu Alexandros G. Dimakis UT Austin dimakis@austin.utexas.edu Sewoong Oh University of Washington sewoong@cs.washington.edu
Pseudocode No The paper describes the algorithm steps in text but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Github Repo: https://github.com/revbucket/dual-verification
Open Datasets Yes We apply Zono Dual to a variety of networks trained on MNIST and CIFAR-10
Dataset Splits Yes Figure 1 (left) contains the distribution of reported bounds over every correctly-classified example in the MNIST validation set.
Hardware Specification No The paper states it is 'amenable to GPU acceleration' but does not specify exact GPU models, CPU types, or other detailed hardware specifications in the main text.
Software Dependencies Yes Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2021. URL https://www. gurobi.com.
Experiment Setup Yes We then decompose each zonotope into 2-dimensional partitions, using the similarity heuristic for fully-connected networks, and the spatial heuristic for convolutional networks. Then we perform 1000 iterations of the Adam optimizer during the dual ascent procedure.