[论文解读] Deep Comprehensive Correlation Mining for Image Clustering

Jianlong Wu, Keyu Long+5 more|arXiv (Cornell University)|Apr 15, 2019

Domain Adaptation and Few-Shot Learning参考文献 58被引用 18

一句话总结

本文提出深度综合相关性挖掘（DCCM），一种端到端的深度聚类框架，通过联合利用样本级相关性、对输入变换的局部鲁棒性以及特征间的三元组互信息，学习具有区分性的表征。DCCM 在 CIFAR-10 上实现了 62.3% 的聚类准确率，相比最先进方法提升了 10.1%。

ABSTRACT

Recent developed deep unsupervised methods allow us to jointly learn representation and cluster unlabelled data. These deep clustering methods mainly focus on the correlation among samples, e.g., selecting high precision pairs to gradually tune the feature representation, which neglects other useful correlations. In this paper, we propose a novel clustering framework, named deep comprehensive correlation mining(DCCM), for exploring and taking full advantage of various kinds of correlations behind the unlabeled data from three aspects: 1) Instead of only using pair-wise information, pseudo-label supervision is proposed to investigate category information and learn discriminative features. 2) The features' robustness to image transformation of input space is fully explored, which benefits the network learning and significantly improves the performance. 3) The triplet mutual information among features is presented for clustering problem to lift the recently discovered instance-level deep mutual information to a triplet-level formation, which further helps to learn more discriminative features. Extensive experiments on several challenging datasets show that our method achieves good performance, e.g., attaining $62.3\%$ clustering accuracy on CIFAR-10, which is $10.1\%$ higher than the state-of-the-art results.

研究动机与目标

• 解决现有深度聚类方法仅关注成对相关性而忽略更高级语义与结构信息的局限性。
• 通过高置信度伪标签引入类别级监督，提升特征的区分性。
• 通过利用小幅度输入扰动下的局部不变性，增强模型鲁棒性。
• 将实例级互信息扩展为三元组级形式，以提升特征学习效果。

提出的方法

• 该方法使用深度卷积神经网络生成预测特征，并通过约束其趋近于 one-hot 分布，以实现可靠的伪标签分配。
• 基于特征间的余弦相似度构建相似性图，并采用高阈值截断策略，选择高置信度的伪图边用于监督。
• 通过训练网络使其对同一张图像的原始版本和变换版本产生一致的预测，实现局部鲁棒性。
• 最大化同一样本的浅层、深层及图增强特征之间的三元组互信息，以保留跨层的区分性信息。
• 框架以端到端方式联合优化三个损失组件：伪标签损失、鲁棒性损失和三元组互信息损失。
• 采用渐进式训练策略，通过多轮迭代逐步优化伪图和伪标签，以提升监督质量。

Figure 1: Comprehensive correlations mining. (a) Various correlations; (b) Connect pair-wise items in higher semantic level progressively; (c) Better results of DCCM than the state-of-the-art DAC [ 9 ] on CIFAR-10 [ 29 ] . Best viewed in color!

实验结果

研究问题

• RQ1 通过伪标签引入类别级信息，是否能超越仅学习成对相关性的方法，进一步提升深度聚类性能？
• RQ2 对输入变换的局部鲁棒性约束，对无监督聚类中特征表示质量有何影响？
• RQ3 将实例级互信息扩展为三元组级形式，是否能进一步增强特征的区分性？
• RQ4 样本相关性、变换鲁棒性以及层间相关性等综合相关性，共同作用下对聚类准确率的提升程度如何？

主要发现

• DCCM 在 CIFAR-10 上实现了 62.3% 的聚类准确率，相比之前最先进方法提升了 10.1%。
• 引入伪标签监督后，聚类准确率相比仅使用伪图的方法绝对提升了 7.1%。
• 与仅使用伪图和伪标签的基线相比，三元组互信息损失额外带来了 4.0% 的准确率提升。
• BCubed 精确率与召回率曲线显示，伪图质量随训练轮次逐步提升，表明自监督机制有效。
• 最大预测概率的分布显示，初始时仅有 <10% 的样本置信度 >0.7，而收敛时接近 80%，证实了学习特征的 one-hot 行为。
• 选择伪标签的高阈值（0.95）至关重要，性能随阈值提高而上升，表明低置信度标签会损害学习效果。

Figure 2: The pipeline of the proposed DCCM method. Based on the ideally one-hot prediction feature, we compute the highly-confident pseudo-graph and pseudo-label to guide the feature learning of both original and transformed samples, investigating both correlations among different samples and local