Deep learning application component interpretable method based on feature map and class activation mapping

Abstract

The invention discloses a deep learning application component interpretable method based on a feature map and class activation mapping. The method is characterized in that information such as input ofa deep learning model, a static structure of the model and a dynamic reasoning process of the model is visually displayed on the basis of a visualization technology around deep learning application evaluation, understanding of a tester on the deep learning technology is enhanced, and then quality evaluation of deep learning application is achieved. According to the method, intuitive understandingis provided for software testers, and detailed information is extracted, rearranged, converted and visualized. The method has the following beneficial effects that for evaluation of deep learning application, after adversarial samples are input, the expression behavior visualization of the model can be embodied. The method is mainly implemented in three aspects of deep learning application component visualization based on a feature map, model reasoning process visualization based on class activation mapping and a deep learning application evaluation method based on visualization.

Description

technical field [0001] The present invention is an instance-based DNN visualization tool. Focusing on deep learning application evaluation, based on visualization technology, visually display information such as deep learning model input, model static structure, model dynamic reasoning process, etc., strengthen testers' understanding of deep learning technology, and then realize the quality evaluation of deep learning applications . Background technique [0002] Deep neural network technology has received increasing attention in the field of software engineering. Deep learning models are used to accomplish various software engineering tasks and are embedded in many software applications. However, analyzing and understanding its behavior is a daunting task for testers. On the one hand, unlike traditional programs that express business logic symbolically, the internal logic of deep learning applications is not easily understood by testers; on the other hand, the testing sta...

AI Technical Summary

Problems solved by technology

Also, during the automated training phase, engineers had little to no interaction with the model as the hyperparameters in the model were constantly being tuned every epoch

But after training, it is possible to interact with the model

Images