Explaining Deep Convolutional Neural Networks via Latent Visual-Semantic Filter Attention
Explaining Deep Convolutional Neural Networks via Latent Visual-Semantic Filter Attention
复制标题
DOI：
10.1109/cvpr52688.2022.00815
复制
发表时间：
2022-04
期刊：
影响因子：
--
通讯作者：
Yu Yang;Seung Wook Kim;Jungseock Joo
中科院分区：
文献类型：
--
作者：
Yu Yang;Seung Wook Kim;Jungseock Joo
可解释性是可视化模型的一个重要属性，它有助于研究者和用户理解复杂模型的内部机制。然而，在没有直接监督的情况下，产生关于学习表征的语义解释是具有挑战性的。我们提出了一个通用框架，La-tent视觉语义解释器（LaViSE），用于教任何现有的卷积神经网络在过滤器级别生成关于其自身潜在表示的文本描述。我们的方法使用通用图像数据集，使用图像和类别名称构建视觉空间和语义空间之间的映射。然后将映射转移到没有语义标签的目标领域。所提出的框架采用模块化结构，能够分析任何训练过的网络，无论其原始训练数据是否可用。我们表明，我们的方法可以在训练数据集中定义的类别集之外为学习过滤器生成新的描述，并在多个数据集上执行广泛的评估。我们还展示了我们的无监督数据集偏差分析方法的新应用，该方法允许我们自动发现数据集中隐藏的偏差或比较不同的子集，而无需使用额外的标签。数据集和代码被公开，以方便进一步的研究。11https://github.com/YuYang0901/LaViSE
Interpretability is an important property for visual mod-els as it helps researchers and users understand the in-ternal mechanism of a complex model. However, gener-ating semantic explanations about the learned representation is challenging without direct supervision to produce such explanations. We propose a general framework, La-tent Visual Semantic Explainer (LaViSE), to teach any ex-isting convolutional neural network to generate text de-scriptions about its own latent representations at the filter level. Our method constructs a mapping between the vi-sual and semantic spaces using generic image datasets, using images and category names. It then transfers the map-ping to the target domain which does not have semantic la-bels. The proposedframework employs a modular structure and enables to analyze any trained network whether or not its original training data is available. We show that our method can generate novel descriptions for learned filters beyond the set of categories defined in the training dataset and perform an extensive evaluation on multiple datasets. We also demonstrate a novel application of our method for unsupervised dataset bias analysis which allows us to auto-matically discover hidden biases in datasets or compare dif-ferent subsets without using additional labels. The dataset and code are made public to facilitate further research.11https://github.com/YuYang0901/LaViSE