Related Works

Deep clustering models.

Existing deep clustering methods can be classified into two main categories: separately and jointly optimization. The separately optimization methods typically first train a feature extractor by self-supervised task such as deep autoencoder models, then traditional clustering methods such as k-means [10], GMM [37] or spectral clustering [1] are applied to obtain the clustering results. There are also some works [27] using density-based clustering algorithm such as DBSCAN [26] to avoid an explicit choice of number of centroids. However, the separately methods require a two-step optimization which lack the ability to train the model in an end-to-end manner to learning representation that is more suitable for clustering. On the contrary, the jointly methods are becoming more popular in the era of deep learning. One prominent approach is the Deep Embedded Clustering (DEC) model [33], which leverages an autoencoder network to map data to a lower-dimensional representations and then optimize the clustering loss KL-divergence between the soft assignments of data to centroids and an adjusted target distribution with concentrated cluster assignments. Deep clustering model (DCN) [35] jointly optmize the dimensionality reduction and k-means clustering objective functions via learning a deep autoencoder and a set of k-means centroids in the embedding space. JULE [36] formulates the joint learning in a recurrent framework, which incorporates agglomerative clustering technique as a forward pass in neural networks. More recently, some works [15, 24, 34] also propose to use contrastive learning by data augmentation techniques to obtain more discriminative representations for downstream clustering tasks.

However, most existing deep clustering methods focus on optimizing a fixed set of centroids, which limits their transferability as they struggle to handle distribution drift between different source and target domains. In contrast, our proposed model takes a different approach by adapting the centroids to learned latent embeddings, allowing it to be aware of distribution drift between domains and enhance its transferability.

Attention models.

Attention models [2, 8, 29, 39] have gained significant attention in the field of deep learning, revolutionizing various tasks across natural language processing, computer vision, and sequence modeling. These works collectively demonstrate the versatility and effectiveness of attention models in capturing informative relationships between data samples.

Deep metric learning.

Our method is also related to deep metric learning methods that aim to learn representations from high-dimensional data in such a way that the similarity or dissimilarity between samples can be accurately measured. One prominent approach is the Contrastive Loss [7], which encourages similar samples to have smaller distances in the embedding space. Siamese networks [5] learns embeddings by comparing pairs of samples and optimizing the contrastive loss. More recently, the Angular Loss [32] incorporates angular margins to enhance the discriminative power of the learned embeddings. Proxy-NCA [21] employs proxy vectors to approximate the intra-class variations, enabling large-scale metric learning.

Connection with Unsupervised Domain Adaption (UDA) methods.

While both our work and existing Unsupervised Domain Adaptation (UDA) methods [6, 16, 18] involve transferring models from source domains to target domains, the primary goal of our paper differs significantly from UDA tasks. UDA methods assume the presence of labeled data in the source domains, allowing the model to be trained in a supervised manner. In contrast, our paper focuses on a scenario where no labels are available in the source domain, necessitating the use of unsupervised learning techniques. This key distinction highlights the unique challenges and approaches we address in our research.