Advancements in Neural Network Verification and Structural Health Monitoring

The field is moving towards developing more robust and efficient methods for verifying neural networks and monitoring structural health. Researchers are exploring new techniques such as Bayesian surrogates, conformal prediction, and branch-and-bound methods to improve the accuracy and reliability of these systems. Additionally, there is a growing interest in applying these methods to real-world problems, such as infrastructure monitoring and maintenance. Noteworthy papers include: Prophecy, which presents a tool for automatically inferring formal properties of feed-forward neural networks. BaB-prob, which proposes a branch-and-bound framework for probabilistic verification of neural networks. Bayesian Surrogates for Risk-Aware Pre-Assessment of Aging Bridge Portfolios, which demonstrates the effectiveness of Bayesian neural network surrogates for rapid structural pre-assessment of bridge portfolios.

Sources

Prophecy: Inferring Formal Properties from Neuron Activations

Real Time Fatigue Crack Growth Monitoring Using High Precision Control and Data Acquisition Systems

Bayesian Surrogates for Risk-Aware Pre-Assessment of Aging Bridge Portfolios

Conformal Prediction for Signal Temporal Logic Inference

BaB-prob: Branch and Bound with Preactivation Splitting for Probabilistic Verification of Neural Networks

Development and Field Validation of a Fully Customised Vehicle Scanning System on Two Full-Scale Bridges

On Integer Programming for the Binarized Neural Network Verification Problem

Built with on top of