Research on Automatic Generation and Mutation Method of Neural Network Based on SMT

This study proposes a neural network automatic construction and mutation model based on statistical machine translation constraints, in an effort to improve the efficiency of neural network design and the level of optimization accuracy. Relying on the neural network calculation graph construction technology, SMT constraint modeling is implemented, and a set of network structure automatic construction algorithm systems are developed. The mutation technology is used to optimize the existing network and implement performance testing. This experiment uses AFL++ and gcov tools to perform fuzz testing and coverage detection on the constructed ONNX model, and to perform performance detection and effect analysis on the automatically generated model and its variants. The experimental results confirm that the model based on automation shows obvious superiority in coverage testing compared with traditional methods, especially in the sensitive interface of edge logic to the compiler triggering effect. The mutation strategy significantly improves the comprehensiveness of the test. Empirical analysis shows that model diversity has a significant positive impact on the improvement of optimization performance.