Classifier visual interpretation stylex: Google, MIT, etc. have found the key attributes that affect image classification

Neural networks can perform various tasks very well , But how do they make decisions ？ for example , In the image classification task , How does the model determine that an image belongs to one category rather than another , This is usually a puzzle . Explain how neural networks make decisions , May have a significant social impact in some areas , For example, medical image analysis and automatic driving .

Previous methods of visual interpretation of classifiers （ For example, attention maps Grad-CAM））, It will emphasize which areas in the image have an impact on the classification , But these methods do not explain which attributes in these areas determine the classification results ： For example, color 、 shape , Or are other attributes that determine the model for such classification . Another method is to provide interpretation by smoothly converting images between one class and another （ for example GANalyze）. However , These methods tend to change all attributes at once （ Such as color 、 shape ）, therefore , It's hard to isolate the impact of a single attribute .

last year , From Google 、 Hebrew University 、 MIT Researchers from other institutions have proposed a new visual interpretation method of classifier , Related papers were published by ICCV 2021 receive . They proposed StylEx , It can automatically discover and visualize the decoupling attributes that affect the classifier （disentangled attributes）.StylEx Allows you to explore the impact of a single attribute by manipulating these attributes individually （ in other words , Changing one property does not affect other properties ）.StylEx Applicable to a wide range of fields , Including animals 、 Leaf 、 Facial and retinal images . The results of this study show that ,StylEx The attributes found match the semantic attributes very well , A meaningful image specific interpretation can be generated , And it can be explained by people in User Research .

• Address of thesis ：https://arxiv.org/pdf/2104.13369.pdf

• Thesis home page ：https://explaining-in-style.github.io/