Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Parallel techniques for neural network verification

Monogyios Antonios

Full record

---

Summary

Parallel Neural Networks verifiers are software tools that leverage parallel architectures to prove fast and rigorously the correct functionality of neural networks in mission-critical environments such as aviation, autonomous driving, etc. In this thesis, we present a plug-in based on Bayesian Optimization that uses transfer learning between different safety specifications to accelerate their parallel verification, while also providing an alternative method based on the state of neurons inside the neural network under verification.In the proposed method of transferring between safety specifications, the plug-in utilizing the available computational resources receives as input a trained neural network and a safety specification applied on its input dimensions. Accordingly, performing a few verification trials it creates a Bayesian Optimization model, which can accurately predict the verification time for any parallelization scheme (input split). Thus, it can provide a parallelization scheme minimizing the overall verification time. Then, it utilizes the trained model to predict parallelization schemes that minimize the overall execution time for new safety specifications that fully or partially overlap with the original specification.The second operational mode is based on performing few prediction trials with the neural network under verification instead of verification. The process creates and trains a Bayesian Optimization model that can accurately predict the number of active and inactive ReLU neurons for a parallelization scheme. As a result, it can provide a parallelization scheme that minimizes the number of unstable neurons, since minimizing this number is accompanied by a minimization on the verification time.Summarizing, we perform an experimental evaluation of the 2 methods of operation in our proposed plug-in, evaluating the quality of the proposed parallelization schemes based on their ranking relative to every other, and provide the results of our measurements. The results gathered during the experimental evaluation prove the efficiency of our methods in regard to accelerating the parallel verification of neural networks.

Available Files

Services