Calculation graph evolution AI model automatic generation method based on structural similarity

Abstract

The invention provides a calculation graph evolution AI model automatic generation method based on structural similarity. In the AI model automatic generation method based on calculation graph evolution, a graph similarity technology, namely a method of combining multiple graph similarity calculations, is adopted for calculating the similarity degree between a generated new network model and a known network model in the automatic generation process of an AI model, and a large number of repeated models, similar models and performance similar models are inhibited from appearing within a certainsimilarity threshold range. The diversity of model samples in the model search process can be effectively ensured, and the model network search success rate is improved. The performance degradation ofthe search network is obviously reduced. Meanwhile, the similarity threshold is dynamically adjusted according to the search efficiency, the function of jumping out of local optimum can be achieved according to model samples, and model search is accelerated. Therefore, the automatic generation efficiency of the AI model is improved.

Description

technical field [0001] The invention relates to the related technical field of AI models (AI models, namely artificial intelligence models), in particular to a method for automatically generating AI models based on structural similarity calculation graph evolution. Background technique [0002] Automatic generation of AI models is a cutting-edge research field. Automatic model generation can generate simpler and more efficient neural network models according to the distribution of data. The search space automatically generated by the AI ​​model is f n ×2 n(n-1) / 2 , where f is the number of different neuron operators, and n is the maximum depth of the neural network. It can be seen that during the generation process, with the increase of supported neural network operators and the deepening of the network model, the complexity of the problem may become a problem that approaches an infinite search space, resulting in the inability to solve it. [0003] At present, the main ...

AI Technical Summary

Problems solved by technology

[0004] In the process of network model search and design, there will be a large number of repeated network models with similar structures or similar performance. Because the calculation process of model performance is time-consuming, a large number of repeated structures and similar models will cause the diversity of samples in the evolutionary algorithm process. Reduced performance, resulting in low search efficiency of the entire evolutionary algorithm, poor performance of the searched model, etc.

Images