Towards Visually Explaining Variational Autoencoders
Towards Visually Explaining Variational Autoencoders
复制标题
DOI：
10.1109/cvpr42600.2020.00867
复制
发表时间：
2019-11
期刊：
影响因子：
--
通讯作者：
Wenqian Liu;Runze Li;Meng Zheng;S. Karanam;Ziyan Wu;B. Bhanu;R. Radke;O. Camps
中科院分区：
文献类型：
--
作者：
Wenqian Liu;Runze Li;Meng Zheng;S. Karanam;Ziyan Wu;B. Bhanu;R. Radke;O. Camps
卷积神经网络（CNN）模型可解释性的最新进展在可视化和理解模型预测方面取得了令人印象深刻的进展。特别是，基于梯度的视觉注意方法最近推动了许多使用视觉注意图作为视觉解释手段的努力。然而，一个关键问题是，这些方法是为分类和分类任务而设计的，并且它们扩展到解释生成模型，例如变分自编码器（VAE）不是微不足道的。在这项工作中，我们朝着弥合这一关键差距迈出了一步，提出了第一种通过基于梯度的注意力来直观地解释VAEs的技术。我们提出了从学习的潜在空间产生视觉注意的方法，并证明了这种注意解释不仅仅是解释VAE预测。我们展示了如何使用这些注意力图来定位图像中的异常，并在MVTec-AD数据集上展示了最先进的性能。我们还展示了如何将它们注入模型训练中，帮助引导VAE学习改进的潜在空间解纠缠，在Dsprites数据集上进行了演示。
Recent advances in Convolutional Neural Network (CNN) model interpretability have led to impressive progress in visualizing and understanding model predictions. In particular, gradient-based visual attention methods have driven much recent effort in using visual attention maps as a means for visual explanations. A key problem, however, is these methods are designed for classification and categorization tasks, and their extension to explaining generative models, e.g., variational autoencoders (VAE) is not trivial. In this work, we take a step towards bridging this crucial gap, proposing the first technique to visually explain VAEs by means of gradient-based attention. We present methods to generate visual attention from the learned latent space, and also demonstrate such attention explanations serve more than just explaining VAE predictions. We show how these attention maps can be used to localize anomalies in images, demonstrating state-of-the-art performance on the MVTec-AD dataset. We also show how they can be infused into model training, helping bootstrap the VAE into learning improved latent space disentanglement, demonstrated on the Dsprites dataset.