Heterogeneous Tri-stream Clustering Network

Abstract

Contrastive deep clustering has recently gained significant attention with its ability of joint contrastive learning and clustering via deep neural networks. Despite the rapid progress, previous works mostly require both positive and negative sample pairs for contrastive clustering, which rely on a relative large batch-size. Moreover, they typically adopt a two-stream architecture with two augmented views, which overlook the possibility and potential benefits of multi-stream architectures (especially with heterogeneous or hybrid networks). In light of this, this paper presents a new end-to-end deep clustering approach termed Heterogeneous Tri-stream Clustering Network (HTCN). The tri-stream architecture in HTCN consists of three main components, including two weight-sharing online networks and a target network, where the parameters of the target network are the exponential moving average of that of the online networks. Notably, the two online networks are trained by simultaneously (i) predicting the instance representations of the target network and (ii) enforcing the consistency between the cluster representations of the target network and that of the two online networks. Experimental results on four challenging image datasets demonstrate the superiority of HTCN over the state-of-the-art deep clustering approaches. The code is available at https://github.com/dengxiaozhi/HTCN.

Deep Multiview Clustering via Iteratively Self-Supervised Universal and Specific Space Learning

“…Specifically, we select four nondeep clustering baseline methods for comparison, including K-means [38], SC [19], AC [20] and NMF [21]. Besides, deep clustering baseline methods involves AE [22], DAE [23], DCGAN [39], DeCNN [40], VAE [41], JULE [25], DEC [24], DAC [37], DCCM [42], PICA [26], IDFD [8], CC [7], ProPos [12], HTCN [13] and SACC [10]. We have reproduced the results of IDFD, CC and ProPos by using their official code and strictly following the suggested settings.…”

“…Recently, deep clustering has made significant progress with the assistance of contrastive learning. Some deep clustering methods based on contrastive learning have achieved remarkable performance [7], [8], [10], [12], [13]. Specifically, [8] presented the IDFD method to learn similarities among instances and reduce correlations within features by adopting the idea of instance discrimination [27] and spectral clustering [19].…”

“…In terms of data augmentations, we take the same augmentations as SimCLR [1], which consists of ResizedCrop, ColorJitter, Grayscale and HorizontalFlip. However, GaussianBlur has been removed since we adopt only a small image size for all datasets [7], [10], [13]. In terms of mini-batch, the batch size is fixed at 256 for CIFAR-10 and CIFAR-100, and 128 for the rest of the datasets.…”

“…Subsequently, several excellent deep clustering works have emerged building upon the foundation of CC. Especially, [13] expanded the dual-branch structure of CC to a triple-branch structure and proposed a tri-stream clustering network called HTCN. Besides, the work presented in the SACC [10] explored the impact of strong and weak augmentations on contrastive clustering.…”

“…E-mail: reganzhx@gmail.com, huangdonghere@gmail.com. [11]- [13]. Specifically, they typically adopt an asymmetric network architecture to perform representation learning and discriminative clustering simultaneously.…”