Explaining in Style: Training a GAN to explain a classifier in StyleSpace

Abstract

Image classification models can depend on multiple different semantic attributes of the image. An explanation of the decision of the classifier needs to both discover and visualize these properties. Here we present StylEx, a method for doing this, by training a generative model to specifically explain multiple attributes that underlie classifier decisions. A natural source for such attributes is the StyleSpace of StyleGAN, which is known to generate semantically meaningful dimensions in the image. However, because standard GAN training is not dependent on the classifier, it may not represent these attributes which are important for the classifier decision, and the dimensions of StyleSpace may represent irrelevant attributes. To overcome this, we propose a training procedure for a StyleGAN, which incorporates the classifier model, in order to learn a classifier-specific StyleSpace. Explanatory * indicates equal contributions; Work performed by authors while working at Google. attributes are then selected from this space. These can be used to visualize the effect of changing multiple attributes per image, thus providing image-specific explanations. We apply StylEx to multiple domains, including animals, leaves, faces and retinal images. For these, we show how an image can be modified in different ways to change its classifier output. Our results show that the method finds attributes that align well with semantic ones, generate meaningful image-specific explanations, and are human-interpretable as measured in user-studies. 1

LARGE: Latent-Based Regression through GAN Semantics

“…Indeed, for some attributes, even finding a disentangled latent direction is infeasible. Furthermore, similar to other methods which rely on StyleGAN [7,29], our method obtains better results when operating on images within the domain used to train the GAN. This limitation stems in part from to the inability of current GAN Inversion methods to reconstruct out-of-domain images while preserving latent semantics.…”

“…In the context of discriminative tasks, several recent methods have proposed to utilize GANs for additional purposes. Lang et al [29] used StyleGAN [27] to visualize counterfactual examples for explaining a pretrained classifier's predictions. Chai et al [7] used style-mixing in the fine-layers of StyleGAN to generate augmentations that are ensembled together at test-time.…”

Self-Distilled StyleGAN: Towards Generation from Internet Photos

“…In the first step, we jointly train a generator G and an encoder E. We use the popular Style-GAN2 generator [16], and train it to produce realistic images using the original objective as have been employed by Karras et al [12,15,16], denoted L GAN . The encoder is trained to reconstruct both real and synthesized images, similar to the work of Lang et al [19]. Let G n (z) be an unconditionally generated image from normally distributed noise vector z, we denote its reconstruction as G(E(G n (z))), where G(w) refers to applying the generator over latent code w (i.e.…”

Which Style Makes Me Attractive? Interpretable Control Discovery and Counterfactual Explanation on StyleGAN

“…GANalyze [8] takes advantages of the GAN-based model to visualize what a CNN model learns about high-level cognitive properties. StylEx [19] proposes to incorporate the classifier into the training process of StyleGAN and learn a classifier-specific StyleSpace. Sauer and Geiger [26] propose to disentangle object shape, object texture and background in the image generation process and generate structured conterfacturals which help improve the robustness and interpretability of classifiers.…”

“…On one hand, it is difficult to sample valid counterfactuals from high-dimensional image manifold by simply altering pixels. To address this difficulty, some recent studies consider to exploit image generation techniques to produce counterfactuals [7,19,30]. On the other hand, even though image counterfactuals can be sampled using generators, it is sometimes still difficult to explain what attributes or concepts are altered.…”