Google AI Introduces ‘StylEx’: A New Approach For A Visual Explanation Of Classifiers

Neural networks are capable of completing a wide range of tasks. Understanding how they arrive at their decisions, on the other hand, is frequently a mystery that goes unexplored. Explaining a neural model’s decision process could have a significant social impact in domains where human oversight is crucial, like medical image processing and autonomous driving. These revelations could be instrumental in advising healthcare practitioners and possibly enhancing scientific breakthroughs.

For a visual explanation of classifiers, there have been approaches such as attention maps. They show which parts of an image have an impact on classification. They are, however, unable to explain how the attributes within those zones affect the classification conclusion. Other methods demonstrate by gracefully transitioning the image from one class to the next. They continue to fall short in isolating the various influencing factors.

Researchers at Google AI unveil a new technique for a visual explanation of classifiers called StylEx. StylEx finds and visualizes disentangled attributes that affect a classifier automatically. As a result, it is possible to investigate the impact of various traits by modifying them separately. Changes in one characteristic do not influence the other attributes, which is a significant benefit of this strategy. Animals, foliage, faces, and retinal pictures are just a few of the areas where StylEx can be used. StylEx finds qualities that correspond well with semantic ones and generates meaningful image-specific explanations, according to the research.

How does StylEX work given a classifier and input image?

To generate high-quality images, the StyleGAN2 architecture is used. It is carried out in two stages:

Phase 1: Training StylEx

“StyleSpace” is a disentangled latent space in StyleGAN2. Individual semantically significant properties of the images in the training dataset are stored in this area. StyleGAN training, on the other hand, is not dependent on the classifier because it may not represent the essential features that determine the classifier. To satisfy the classifier, the researchers choose to train a StyleGAN-like generator. This encourages StyleSpace to accommodate classifier-specific features in addition to resolving the issue. This is accomplished by adding two more components to the StyleGAN generator:

  • Encoder
  • GAN

With a reconstruction loss, they are trained jointly. Hence, the created output resembles the input image in appearance. For the same reason, the generator can be used on any image as an input. However, the image’s visual similarity isn’t up to par, as it may miss small optical characteristics that are crucial to a particular classifier. The researchers modified the StyleGAN training to include a classification-loss component. This additional component compels the output image’s classifier probability to be the same as its classifier probability. Hence, the resulting image contains small visual characteristics crucial to the classifier.

Phase 2: Now, the disentangled attributes have to be extracted

Following the training phase, the StyleSpace is combed for attributes that substantially impact the classifier. Each StyleSpace coordinate is changed, and its effect on the categorization probability is measured to complete the task. The top qualities for a given image are chosen to maximize the change in classification likelihood. By repeating the method on vast data collections, the top-K class-specific attributes can be determined.

Both binary and multi-class classifiers can be used using the method described above. The top attributes found by Google AI’s algorithm conform to coherent semantic conceptions when interpreted by humans in all domains tested, as confirmed by human evaluation.

The path ahead:

Finally, the researchers’ strategy allows for meaningful explanations for a given classifier on a specific image or class. According to Google’s AI Principles, the researchers believe their technique is a promising step toward detecting and mitigating previously undisclosed biases in classifiers and datasets. Furthermore, shifting the focus to multiple-attribute-based explanation is critical for providing fresh insights into previously opaque classification methods and assisting in the scientific discovery process.

Paper: https://arxiv.org/pdf/2104.13369.pdf

Project: https://explaining-in-style.github.io/

Github: https://github.com/google/explaining-in-style

Reference: https://ai.googleblog.com/2022/01/introducing-stylex-new-approach-for.html