Researchers released NNV3, the latest version of the Neural Network Verification (NNV) tool, a MATLAB framework for formal verification of deep learning models and learning-enabled cyber-physical systems. NNV3 adds new Star-set verifiers for weight perturbation, video/3D inputs, and graph neural networks, plus probabilistic reachability and fairness certification. The release ships with new benchmarks and a unified documentation site, broadening the tool's applicability to AI systems practitioners need to verify.
An arXiv paper introduces GNNV, an extension of the neural network verification (NNV) framework to graph-structured inputs via GraphStar sets, which capture uncertainty over node and edge features. It enables sound verification of GCN and GINE layers and is evaluated on power system tasks and graph classification benchmarks. The method reportedly delivers tighter robustness guarantees than CORA and the first edge-aware guarantees for GINE-based power flow models.
An arXiv paper introduces GNNV, an extension of the neural network verification (NNV) framework to graph-structured inputs via GraphStar sets, which capture uncertainty over node and edge features. It enables sound verification of GCN and GINE layers and is evaluated on power system tasks and graph classification benchmarks. The method reportedly delivers tighter robustness guarantees than CORA and the first edge-aware guarantees for GINE-based power flow models.
Researchers introduce an inprocessing framework for neural network verification that uses lookahead to derive lemmas over unstable ReLU phases, collected into an implication graph to prune search and vivify boolean cuts. Instantiating it in Marabou and α-β-CROWN improves performance, proving up to 34% more instances unsatisfiable.
Researchers introduce an inprocessing framework for neural network verification that uses lookahead to derive lemmas over unstable ReLU phases, collected into an implication graph to prune search and vivify boolean cuts. Instantiating it in Marabou and α-β-CROWN improves performance, proving up to 34% more instances unsatisfiable.
A new arXiv paper introduces an approach to improve Branch and Bound (BaB) verification of neural networks by mining verdict boundaries across the splitting tree, rather than solving each subproblem sequentially. This could reduce the computational cost of complete verification.
A new arXiv paper introduces an approach to improve Branch and Bound (BaB) verification of neural networks by mining verdict boundaries across the splitting tree, rather than solving each subproblem sequentially. This could reduce the computational cost of complete verification.
A new arXiv paper introduces Constrained Lagrangian Abstract Domain (CLAD), a neural network verification method that handles input regions defined by combinations of convex constraints, unlike existing domains restricted to Lp-norm balls. CLAD matches state-of-the-art tools on unconstrained properties and verifies substantially more constrained-property instances. This matters to practitioners working on formal guarantees for networks under realistic, structured perturbations.
A new arXiv paper introduces Constrained Lagrangian Abstract Domain (CLAD), a neural network verification method that handles input regions defined by combinations of convex constraints, unlike existing domains restricted to Lp-norm balls. CLAD matches state-of-the-art tools on unconstrained properties and verifies substantially more constrained-property instances. This matters to practitioners working on formal guarantees for networks under realistic, structured perturbations.
A new software tool, described in an arXiv paper, automates numerical stability analysis of deep learning operators using CESTAC. It detects instability sources and monitors stability during training and inference in a single computation pass, addressing finite-precision errors.
A new software tool, described in an arXiv paper, automates numerical stability analysis of deep learning operators using CESTAC. It detects instability sources and monitors stability during training and inference in a single computation pass, addressing finite-precision errors.