Conclusion

We believe that StylEx is a promising step towardsdetection and mitigation of previously unknown biases inclassifiers.

83.2%

In general for all coordinates extracted by StylEx,except one, there is a common word shared by all descrip-tions, and only in two coordinates the most common word isthe same. On the other hand, for Wu et al. less than half ofthe coordinates have a common word in all their descriptions,and two pairs of coordinates have the same most commonword.

Users are shown four animated GIFs, each correspondingto modifying an attribute for a given image.

We first demonstrate that each StylEx attribute corre-sponds to clear visual concepts, and then show that these canbe used to explain classifier outputs on specific images

However, the three followingcriteria seem key to any such method

ee Sec. 4.2.2 for user studiesthat further demonstrate this

Inother words, the modified classifier output may reflect biasesin classifier training, and not a true correlation between thelabel and visual attributes

high accuracy of at least 95% on their test sets.

greedy search (i.e., at each step find the next most influentialattribute for this image, given the subset of attributes selectedso far; halt once the classifier has flipped its classification).We can then visualize the effect of modifying this subset.We refer to this as Subset selection

The simplest is to iterate over StylExattributes, calculate the effect of changing each on the classi-fier output for this image, and return the top-k of these. Wecan then visualize the resulting k modified images. We referto this strategy as Independent selection.

AttFind proceeds as follows: At each iteration it consid-ers all K style coordinates and calculates their effect on theprobability of y.3 4 It then selects the coordinate with largesteffect, and removes all images where changing this coordi-nate had a large effect on their probability to belong to classy

AttFind takes as input the trained model and a set of N im-ages whose predicted label by C is different from y. Foreach class y (e.g., y=“cat” or y=“dog”), AttFind then findsa set Sy of M style coordinates (i.e., Sy ⊂ [1, . . . , K] and|Sy|= M ), such that changing these coordinates increasesthe average probability of the class y on these images.

StylEx architecture.

ablation

LPIPS

The basic GAN trainingrecipe is to train the generator G and an adversarial discrimi-nator D simultaneously [8]. Additionally, we jointly train anencoder E with the generator G, using a reconstruction loss(i.e., the Encoder and Generator function together as an au-toencoder). Finally, we add two components that introducethe classifier into the training procedure.

latent vector w

he observationrecently made by [35] that StyleGAN2 tends to inherentlycontain a disentangled StyleSpace space, which can be usedto extract individual attributes. Thus, we argue that modify-ing coordinates of StyleSpace is a natural approach to ourproblem of modifying classifier-related attributes.

a) A conditionalgenerative model that maps an embedding w into an outputimage. b) An encoder that maps any input image into anembedding w, so that the generator can modify attributesin real images. c) A mechanism for “intervening” with thegeneration process to change visual attributes in the image

cannot visualize/explain well attributes that are not spatiallylocalized, like size, color, etc. In addition, they can showwhich areas of the image may be changed in order to affectthe classification, but not how they should be changed

It should be emphasized that our goal is not to explainthe true label, but rather what classifiers are learning.

Our contributions are as follows

Adversarialexamples [9, 12] , which are slight modifications to the inputxthat change the classification to the wrong class,

A keyadvantage of this approach is that it provides per-exampleexplanations, pinpointing which parts of the input are salienttowards the classification and also how they can be changedin order to obtain an alternative outcome.

A counterfactual explanation is a statement of the form“Had the input xbeen ̃xthen the classifier output would havebeen ̃y instead of y”, where the difference between xand ̃xis easy to explain.

counterfactual example,i.e., visualizing how manipulation of this attribute (style coordinate) affects the classifier output probability.

Deep net classifiers are often described as “black boxes”whose decisions are opaque and hard for humans to under-stand. This significantly holds back their usage, especiallyin areas of high societal impact, such as medical imagingand autonomous driving, where human oversight is critical.

driving latent attributes in the GAN’s StyleSpace to captureclassifier-specific attributes

StylEx, explains the decisions of aclassifier by discovering and visualizing multiple attributes that affect its prediction.