Multiple clustering has gathered significant attention in recent years due to its potential to reveal multiple hidden structures of the data from different perspectives. Most of multiple clustering methods first derive feature representations by controlling the dissimilarity among them, subsequently employing traditional clustering methods (e.g., k-means) to achieve the final multiple clustering outcomes. However, the learned feature representations can exhibit a weak relevance to the ultimate goal of distinct clustering. Moreover, these features are often not explicitly learned for the purpose of clustering. Therefore, in this paper, we propose a novel Dual-Disentangled deep Multiple Clustering method named DDMC by learning disentangled representations. Specifically, DDMC is achieved by a variational Expectation-Maximization (EM) framework. In the E-step, the disentanglement learning module employs coarse-grained and fine-grained disentangled representations to obtain a more diverse set of latent factors from the data. In the M-step, the cluster assignment module utilizes a cluster objective function to augment the effectiveness of the cluster output. Our extensive experiments demonstrate that DDMC consistently outperforms state-of-the-art methods across seven commonly used tasks. Our code is available at https://github.com/Alexander-Yao/DDMC.
Vision-language pre-training methods, e.g., CLIP, demonstrate an impressive zero-shot performance on visual categorizations with the class proxy from the text embedding of the class name. However, the modality gap between the text and vision space can result in a sub-optimal performance. We theoretically show that the gap cannot be reduced sufficiently by minimizing the contrastive loss in CLIP and the optimal proxy for vision tasks may reside only in the vision space. Therefore, given unlabeled target vision data, we propose to learn the vision proxy directly with the help from the text proxy for zero-shot transfer. Moreover, according to our theoretical analysis, strategies are developed to further refine the pseudo label obtained by the text proxy to facilitate the intra-modal proxy learning (InMaP) for vision. Experiments on extensive downstream tasks confirm the effectiveness and efficiency of our proposal. Concretely, InMaP can obtain the vision proxy within one minute on a single GPU while improving the zero-shot accuracy from $77.02\%$ to $80.21\%$ on ImageNet with ViT-L/14@336 pre-trained by CLIP. Code is available at \url{https://github.com/idstcv/InMaP}.
Clustering aims to group similar objects together while separating dissimilar ones apart. Thereafter, structures hidden in data can be identified to help understand data in an unsupervised manner. Traditional clustering methods such as k-means provide only a single clustering for one data set. Deep clustering methods such as auto-encoder based clustering methods have shown a better performance, but still provide a single clustering. However, a given dataset might have multiple clustering structures and each represents a unique perspective of the data. Therefore, some multiple clustering methods have been developed to discover multiple independent structures hidden in data. Although deep multiple clustering methods provide better performance, how to efficiently capture the alternative perspectives in data is still a problem. In this paper, we propose AugDMC, a novel data Augmentation guided Deep Multiple Clustering method, to tackle the challenge. Specifically, AugDMC leverages data augmentations to automatically extract features related to a certain aspect of the data using a self-supervised prototype-based representation learning, where different aspects of the data can be preserved under different data augmentations. Moreover, a stable optimization strategy is proposed to alleviate the unstable problem from different augmentations. Thereafter, multiple clusterings based on different aspects of the data can be obtained. Experimental results on three real-world datasets compared with state-of-the-art methods validate the effectiveness of the proposed method.
Sepsis is a syndrome that develops in response to the presence of infection. It is characterized by severe organ dysfunction and is one of the leading causes of mortality in Intensive Care Units (ICUs) worldwide. These complications can be reduced through early application of antibiotics, hence the ability to anticipate the onset of sepsis early is crucial to the survival and well-being of patients. Current machine learning algorithms deployed inside medical infrastructures have demonstrated poor performance and are insufficient for anticipating sepsis onset early. In recent years, deep learning methodologies have been proposed to predict sepsis, but some fail to capture the time of onset (e.g., classifying patients' entire visits as developing sepsis or not) and others are unrealistic to be deployed into medical facilities (e.g., creating training instances using a fixed time to onset where the time of onset needs to be known apriori). Therefore, in this paper, we first propose a novel but realistic prediction framework that predicts each morning whether sepsis onset will occur within the next 24 hours using data collected at night, when patient-provider ratios are higher due to cross-coverage resulting in limited observation to each patient. However, as we increase the prediction rate into daily, the number of negative instances will increase while that of positive ones remain the same. Thereafter, we have a severe class imbalance problem, making a machine learning model hard to capture rare sepsis cases. To address this problem, we propose to do nightly profile representation learning (NPRL) for each patient. We prove that NPRL can theoretically alleviate the rare event problem. Our empirical study using data from a level-1 trauma center further demonstrates the effectiveness of our proposal.
Sepsis is a life-threatening organ malfunction caused by the host's inability to fight infection, which can lead to death without proper and immediate treatment. Therefore, early diagnosis and medical treatment of sepsis in critically ill populations at high risk for sepsis and sepsis-associated mortality are vital to providing the patient with rapid therapy. Studies show that advancing sepsis detection by 6 hours leads to earlier administration of antibiotics, which is associated with improved mortality. However, clinical scores like Sequential Organ Failure Assessment (SOFA) are not applicable for early prediction, while machine learning algorithms can help capture the progressing pattern for early prediction. Therefore, we aim to develop a machine learning algorithm that predicts sepsis onset 6 hours before it is suspected clinically. Although some machine learning algorithms have been applied to sepsis prediction, many of them did not consider the fact that six hours is not a small gap. To overcome this big gap challenge, we explore a multi-subset approach in which the likelihood of sepsis occurring earlier than 6 hours is output from a previous subset and feed to the target subset as additional features. Moreover, we use the hourly sampled data like vital signs in an observation window to derive a temporal change trend to further assist, which however is often ignored by previous studies. Our empirical study shows that both the multi-subset approach to alleviating the 6-hour gap and the added temporal trend features can help improve the performance of sepsis-related early prediction.
Fine-tuning a pre-trained model can leverage the semantic information from large-scale pre-training data and mitigate the over-fitting problem on downstream tasks with limited training examples. While the problem of catastrophic forgetting in backbone has been extensively studied, the potential bias existing in a pre-trained model due to the corresponding pre-training task and data, attracts less attention. In this work, we investigate this problem by demonstrating that the obtained classifier after fine-tuning will be close to that induced by the pre-trained model. To reduce the bias in the classifier effectively, we introduce a reference distribution obtained from a fixed text classifier, which can help regularize the learned vision classifier. The proposed method, Text Supervised fine-tuning (TeS), is evaluated with diverse pre-trained vision models including ResNet and ViT, and text encoders including BERT and CLIP, on 11 downstream tasks. The consistent improvement with a clear margin over distinct scenarios confirms the effectiveness of our proposal.
Cluster discrimination is an effective pretext task for unsupervised representation learning, which often consists of two phases: clustering and discrimination. Clustering is to assign each instance a pseudo label that will be used to learn representations in discrimination. The main challenge resides in clustering since many prevalent clustering methods (e.g., k-means) have to run in a batch mode that goes multiple iterations over the whole data. Recently, a balanced online clustering method, i.e., SwAV, is proposed for representation learning. However, the assignment is optimized within only a small subset of data, which can be suboptimal. To address these challenges, we first investigate the objective of clustering-based representation learning from the perspective of distance metric learning. Based on this, we propose a novel clustering-based pretext task with online \textbf{Co}nstrained \textbf{K}-m\textbf{e}ans (\textbf{CoKe}) to learn representations and relations between instances simultaneously. Compared with the balanced clustering that each cluster has exactly the same size, we only constrain the minimum size of clusters to flexibly capture the inherent data structure. More importantly, our online assignment method has a theoretical guarantee to approach the global optimum. Finally, two variance reduction strategies are proposed to make the clustering robust for different augmentations. Without keeping representations of instances, the data is accessed in an online mode in CoKe while a single view of instances at each iteration is sufficient to demonstrate a better performance than contrastive learning methods relying on two views. Extensive experiments on ImageNet verify the efficacy of our proposal. Code will be released.
Knowledge distillation is an effective way for model compression in deep learning. Given a large model (i.e., teacher model), it aims to improve the performance of a compact model (i.e., student model) by transferring the information from the teacher. An essential challenge in knowledge distillation is to identify the appropriate information to transfer. In early works, only the final output of the teacher model is used as the soft label to help the training of student models. Recently, the information from intermediate layers is also adopted for better distillation. In this work, we aim to optimize the process of knowledge distillation from the perspective of kernel matrix. The output of each layer in a neural network can be considered as a new feature space generated by applying a kernel function on original images. Hence, we propose to transfer the corresponding kernel matrix (i.e., Gram matrix) from teacher models to student models for distillation. However, the size of the whole kernel matrix is quadratic to the number of examples. To improve the efficiency, we decompose the original kernel matrix with Nystr{\"{o}}m method and then transfer the partial matrix obtained with landmark points, whose size is linear in the number of examples. More importantly, our theoretical analysis shows that the difference between the original kernel matrices of teacher and student can be well bounded by that of their corresponding partial matrices. Finally, a new strategy of generating appropriate landmark points is proposed for better distillation. The empirical study on benchmark data sets demonstrates the effectiveness of the proposed algorithm. Code will be released.
With the development of computational power and techniques for data collection, deep learning demonstrates a superior performance over many existing algorithms on benchmark data sets. Many efforts have been devoted to studying the mechanism of deep learning. One of important observations is that deep learning can learn the discriminative patterns from raw materials directly in a task-dependent manner. It makes the patterns obtained by deep learning outperform hand-crafted features significantly. However, those patterns can be misled by the training task when the target task is different. In this work, we investigate a prevalent problem in real-world applications, where the training set only accesses to the supervised information from superclasses but the target task is defined on fine-grained classes. Each superclass can contain multiple fine-grained classes. In this scenario, fine-grained patterns are essential to classify examples from fine-grained classes while they can be neglected when training only with labels from superclasses. To mitigate the challenge, we propose the algorithm to explore the fine-grained patterns sufficiently without additional supervised information. Besides, our analysis indicates that the performance of learned patterns on the fine-grained classes can be theoretically guaranteed. Finally, an efficient algorithm is developed to reduce the cost of optimization. The experiments on real-world data sets verify that the propose algorithm can significantly improve the performance on the fine-grained classes with information from superclasses only.
With the tremendous success of deep learning in visual tasks, the representations extracted from intermediate layers of learned models, that is, deep features, attract much attention of researchers. The previous analysis shows that those features include appropriate semantic information. By training the deep models on a large-scale benchmark data set (e.g., ImageNet), the features can work well on other tasks. In this work, we investigate this phenomenon and demonstrate that deep features can fail due to the fact that they are learned by minimizing empirical risk. When the distribution of data is different from that of the benchmark data set, the performance of deep features can degrade. Hence, we propose a hierarchically robust optimization to learn more generic features. Considering the example-level and concept-level robustness simultaneously, we formulate the problem as a distributionally robust optimization problem with Wasserstein ambiguity set constraints. An efficient algorithm with the conventional training pipeline is proposed. Experiments on benchmark data sets confirm our claim and demonstrate the effectiveness of the robust deep representations.