Despite the increasing development in recent years of explainability techniques for deep neural networks, only some are dedicated to explaining the decisions made by neural networks for similarity learning. While existing approaches can explain classification models, their adaptation to generate visual similarity explanations is not trivial. Neural architectures devoted to this task learn an embedding that maps similar examples to nearby vectors and non-similar examples to distant vectors in the feature space. In this paper, we propose a agnostic technique that explains the inference of such architectures on a pair of images. The proposed method establishes a relation between the most important features of the abstract feature space and the input feature space (pixels) of an image. For this purpose, we employ a relevance assignment and a perturbation process based on the most influential latent features in the inference. Then, a reconstruction process of the images of the pair is carried out from the perturbed embedding vectors. This process relates the latent features to the original input features. The results indicate that our method produces “continuous” and “selective” explanations. A sharp drop in the value of the function (summarized by a low value of the area under the curve) indicates its superiority over other explainability approaches when identifying features relevant to similarity learning. In addition, we demonstrate that our technique is agnostic to the specific type of similarity model, e.g., we show its applicability in two similarity learning tasks: face recognition and image retrieval.

中文翻译：

---

ExplainLFS：解释从潜在特征空间中的局部扰动进行相似性学习的神经架构

尽管近年来深度神经网络的可解释性技术不断发展，但只有一些技术致力于解释神经网络为相似性学习所做的决策。虽然现有的方法可以解释分类模型，但它们生成视觉相似性解释的适应并不是微不足道的。致力于此任务的神经架构学习一种嵌入，将相似的示例映射到附近的向量，将不相似的示例映射到特征空间中的远处向量。在本文中，我们提出了一种不可知论技术来解释这种架构在一对图像上的推断。所提出的方法建立了抽象特征空间的最重要特征与图像的输入特征空间（像素）之间的关系。为此，我们根据推理中最有影响力的潜在特征采用相关性分配和扰动过程。然后，根据扰动的嵌入向量执行该对图像的重建过程。该过程将潜在特征与原始输入特征相关联。结果表明，我们的方法产生了“连续”和“选择性”解释。函数值的急剧下降（由曲线下面积的低值概括）表明在识别与相似性学习相关的特征时，它比其他可解释性方法具有优越性。此外，我们证明我们的技术与特定类型的相似性模型无关，例如，我们展示了它在两个相似性学习任务中的适用性：人脸识别和图像检索。