Visual Analysis Method for Local Performance of Deep Learning Models Based on Generative Feature Maps

Existing classification model performance evaluation methods mainly focus on the overall performance, but local performance evaluation is very important in applications with high accuracy requirements. Aiming at the challenges of sample data diversity and flexible exploration tasks faced by the local performance evaluation of deep learning models, we propose a visual analysis method based on generative feature maps for the local performance of deep learning models. Firstly, we train a latent space semantic direction discovery model to generate test samples with subtle semantic variations to provide controllable test samples for model evaluation; then, we design a visualization system integrating a direction selector and a performance explorer, in which the direction selector supports the user to generate test samples with specified semantic variations, and the performance explorer consists of a sample graph and a performance graph, where the sample graph displays the distribution of the local samples, and the performance graph visualizes the model’s performance in the local area through colour coding. The performance graph is colour coded to visually display the model’s performance distribution over local samples. Experimental results show that the directions discovered by the proposed latent space semantic direction discovery model outperform the random directions and axes directions in terms of accuracy, interpretability, and naturalness of sample variations; the generative feature maps can effectively identify and explain the performance variations of the classification model on local samples on both MNIST and CIFAR-10 datasets, and identify the local samples that can help to discover the defects of the model samples.