Fast Real-time Counterfactual Explanations
This provides fast, realistic explanations for image classification systems, addressing a bottleneck in interpretability for practical applications, though it is incremental as it builds on existing multi-domain translation methods.
The paper tackles the problem of generating counterfactual explanations for image classification by proposing FRACE, an optimization-free method that uses a transformer trained with GANs to change class-specific semantics while preserving content, achieving state-of-the-art results in quality and speed with real-time inference.
Counterfactual explanations are considered, which is to answer {\it why the prediction is class A but not B.} Different from previous optimization based methods, an optimization-free Fast ReAl-time Counterfactual Explanation (FRACE) algorithm is proposed benefiting from the development of multi-domain image to image translation algorithms. Built from starGAN, a transformer is trained as a residual generator conditional on a classifier constrained under a proposal perturbation loss which maintains the content information of the query image, but just the class-specific semantic information is changed. The transformer can transfer the query image to any counterfactual class, and during inference, our explanation can be generated by it only within a forward time. It is fast and can satisfy the real-time practical application. Because of the adversarial training of GAN, our explanation is also more realistic compared to other counterparts. The experimental results demonstrate that our proposal is better than the existing state of the art in terms of quality and speed.