Research On Interpretability Of Neural Network Model Combined With Semantic Annotation

The interpretability of neural network models is a cutting-edge research direction in the field of artificial intelligence,which is dedicated to exhibiting the internal learning process and decision logic of neural network models in a way that can be understood by humans through interpretable methods.In the field of deep learning,due to the complex hierarchical structure of neural network models,they contain a large amount of code and numerical values that are difficult for humans to understand internally.As a result,their transparency and interpretability cannot meet the basic requirements of humans for trustworthy artificial intelligence systems.Therefore,deep learning is also known as a "black-box" algorithm.Semantic concepts are characteristic attributes that humans can intuitively understand.Research on the interpretability of deep neural network models based on semantic features can intuitively display the internal features and operating principles of neural networks,thereby enhancing human understanding and trust in neural network models.This thesis focuses on improving the accuracy and interpretability of semantic labeling of neural network internal neurons,and based on this,constructing decision relationship logic that runs throughout the entire model,ultimately providing a method for explaining the internal features and decision logic of neural network models from a semantic level.The specific research work includes the following points:1.Semantic consistency analysis of neural network based on model dissection.In response to the limited number of semantic labels in available semantic labeling datasets and the problem of inaccurate semantic labeling of neural units,a semantic-conceptcentered method for neural unit semantic consistency labeling is proposed.By decoupling the representation of neural units for different semantic concepts,multiple neural unit representations for a semantic concept are labeled,resulting in more interpretable semantic concept labeling within the limited dataset.2.Semantic hierarchical reasoning of neural network based on decision tree.A semantic hierarchical reasoning method based on decision tree is proposed to address the problem of the opaque decision logic and difficult to understand decision process of neural network models.By calculating the similarity through neuron weight parameters and feature maps,the connection relationship between neural network layers is constructed according to the inference logic of decision tree.The construction of the whole model’s semantic interpretable decision tree is completed,providing a method based on semantic concept interpretation for explaining the decision logic of neural network models from a global perspective.3.The application of semantic interpretable decision tree.To meet the demand for interpretability of neural network models in real-world scenarios,a system framework and related functional modules are designed.With the use of development frameworks such as Vue and Spring Boot,a semantic interpretable decision tree application system is built and the proposed method for neural network interpretability is applied to the interpretation of specific image recognition processes.