Pretext Task Research Articles

A multi-view contrastive learning for heterogeneous network embedding

Graph contrastive learning has been developed to learn discriminative node representations on homogeneous graphs. However, it is not clear how to augment the heterogeneous graphs without substantially altering the underlying semantics or how to design appropriate pretext tasks to fully capture the rich semantics preserved in heterogeneous information networks (HINs). Moreover, early investigations demonstrate that contrastive learning suffer from sampling bias, whereas conventional debiasing techniques (e.g., hard negative mining) are empirically shown to be inadequate for graph contrastive learning. How to mitigate the sampling bias on heterogeneous graphs is another important yet neglected problem. To address the aforementioned challenges, we propose a novel multi-view heterogeneous graph contrastive learning framework in this paper. We use metapaths, each of which depicts a complementary element of HINs, as the augmentation to generate multiple subgraphs (i.e., multi-views), and propose a novel pretext task to maximize the coherence between each pair of metapath-induced views. Furthermore, we employ a positive sampling strategy to explicitly select hard positives by jointly considering semantics and structures preserved on each metapath view to alleviate the sampling bias. Extensive experiments demonstrate MCL consistently outperforms state-of-the-art baselines on five real-world benchmark datasets and even its supervised counterparts in some settings.

Medical image segmentation based on self-supervised hybrid fusion network.

Automatic segmentation of medical images has been a hot research topic in the field of deep learning in recent years, and achieving accurate segmentation of medical images is conducive to breakthroughs in disease diagnosis, monitoring, and treatment. In medicine, MRI imaging technology is often used to image brain tumors, and further judgment of the tumor area needs to be combined with expert analysis. If the diagnosis can be carried out by computer-aided methods, the efficiency and accuracy will be effectively improved. Therefore, this paper completes the task of brain tumor segmentation by building a self-supervised deep learning network. Specifically, it designs a multi-modal encoder-decoder network based on the extension of the residual network. Aiming at the problem of multi-modal feature extraction, the network introduces a multi-modal hybrid fusion module to fully extract the unique features of each modality and reduce the complexity of the whole framework. In addition, to better learn multi-modal complementary features and improve the robustness of the model, a pretext task to complete the masked area is set, to realize the self-supervised learning of the network. Thus, it can effectively improve the encoder's ability to extract multi-modal features and enhance the noise immunity. Experimental results present that our method is superior to the compared methods on the tested datasets.

Few-shot learning for seismic facies segmentation via prototype learning

The mapping of seismic facies from seismic data is considered a multiclass image semantic segmentation problem. Despite the signification progress made by the deep learning methods in seismic prospecting, the dense prediction problem of seismic facies requires large amounts of annotated seismic facies data, which often are unavailable. These valuable labels are only helpful in one model and field due to geologic heterogeneity. To overcome these challenges, we have developed a few-shot seismic facies segmentation model. Few-shot learning has been designed to learn to perform with very few labels and we design reconstructing masked traces as a pretext task for self-supervised learning to obtain a good feature extractor. By these, this model can use all seismic data from different fields, which is different from image data as the texture-based data. With two different seismic data in turn as a meta-training set and a meta-testing set, our model works well in one- and five-shot settings, which means only one label and five labels, respectively.

INS-GNN: Improving graph imbalance learning with self-supervision

Graph Neural Networks (GNNs) have achieved tremendous success in various applications, such as node classification, link prediction and graph classification. However, graph-structured data is usually imbalanced in many real-world scenarios. When trained on an imbalanced dataset, the performance of GNNs is distant from satisfactory for nodes of minority classes. Due to the small population, these minority nodes have less engagement in the objective function of training and the message passing mechanism behind GNNs exacerbates this problem further, as the information from minority nodes can be overwhelmed by majority nodes in the process of information propagation. To tackle the problems of imbalanced node classification based on GNNs, the most effective way is to promote the engagement of minority nodes during propagation. Therefore, inspired by self-supervised learning on exploring useful information from unlabeled data samples, we propose a novel model-agnostic framework in this paper, named INS-GNN, which tackles the imbalanced node classification problem. Specifically, self-supervised pre-training is firstly used as a pretext task to pre-train the model without label information, which can effectively alleviate the phenomenon that imbalanced labels naturally impose label bias during learning. Then, self-training is exploited to assign pseudo-labels to unlabeled nodes and conduct self-supervised learning to help model training. It can effectively alleviate the problem that obtaining numerous annotation nodes is time-consuming and resource-costly. Finally, self-supervised edge augmentation is utilized to change the structure information of the minority nodes for enhancing the contribution of minority nodes. Experimental results on various real-world datasets demonstrate INS-GNN is able to achieve state-of-the-art performance on solving the imbalanced node classification problem.

Deep fiber clustering: Anatomically informed fiber clustering with self-supervised deep learning for fast and effective tractography parcellation

White matter fiber clustering is an important strategy for white matter parcellation, which enables quantitative analysis of brain connections in health and disease. In combination with expert neuroanatomical labeling, data-driven white matter fiber clustering is a powerful tool for creating atlases that can model white matter anatomy across individuals. While widely used fiber clustering approaches have shown good performance using classical unsupervised machine learning techniques, recent advances in deep learning reveal a promising direction toward fast and effective fiber clustering. In this work, we propose a novel deep learning framework for white matter fiber clustering, Deep Fiber Clustering (DFC), which solves the unsupervised clustering problem as a self-supervised learning task with a domain-specific pretext task to predict pairwise fiber distances. This process learns a high-dimensional embedding feature representation for each fiber, regardless of the order of fiber points reconstructed during tractography. We design a novel network architecture that represents input fibers as point clouds and allows the incorporation of additional sources of input information from gray matter parcellation. Thus, DFC makes use of combined information about white matter fiber geometry and gray matter anatomy to improve the anatomical coherence of fiber clusters. In addition, DFC conducts outlier removal naturally by rejecting fibers with low cluster assignment probability. We evaluate DFC on three independently acquired cohorts, including data from 220 individuals across genders, ages (young and elderly adults), and different health conditions (healthy control and multiple neuropsychiatric disorders). We compare DFC to several state-of-the-art white matter fiber clustering algorithms. Experimental results demonstrate superior performance of DFC in terms of cluster compactness, generalization ability, anatomical coherence, and computational efficiency.

Continuous frame motion sensitive self-supervised collaborative network for video representation learning

Motion, as a feature of video that changes in temporal sequences, is crucial to visual understanding. The powerful video representation and extraction models are typically able to focus attention on motion features in challenging dynamic environments to complete more complex video understanding tasks. However, previous approaches discriminate mainly based on similar features in the spatial or temporal domain, ignoring the interdependence of consecutive video frames. In this paper, we propose the motion sensitive self-supervised collaborative network, a video representation learning framework that exploits a pretext task to assist feature comparison and strengthen the spatiotemporal discrimination power of the model. Specifically, we first propose the motion-aware module, which extracts consecutive motion features from the spatial regions by frame difference. The global–local contrastive module is then introduced, with context and enhanced video snippets being defined as appropriate positive samples for a broader feature similarity comparison. Finally, we introduce the snippet operation prediction module, which further assists contrastive learning to obtain more reliable global semantics by sensing changes in continuous frame features. Experimental results demonstrate that our work can effectively extract robust motion features and achieve competitive performance compared with other state-of-the-art self-supervised methods on downstream action recognition and video retrieval tasks.

Korean Text to Gloss: Self-Supervised Learning approach

Natural Language Processing (NLP) has grown tremendously in recent years. Typically, bilingual, and multilingual translation models have been deployed widely in machine translation and gained vast attention from the research community. On the contrary, few studies have focused on translating between spoken and sign languages, especially non-English languages. Prior works on Sign Language Translation (SLT) have shown that a mid-level sign gloss representation enhances translation performance. Therefore, this study presents a new large-scale Korean sign language dataset, the Museum-Commentary Korean Sign Gloss (MCKSG) dataset, including 3828 pairs of Korean sentences and their corresponding sign glosses used in Museum-Commentary contexts. In addition, we propose a translation framework based on self-supervised learning, where the pretext task is a text-to-text from a Korean sentence to its back-translation versions, then the pre-trained network will be fine-tuned on the MCKSG dataset. Using self-supervised learning help to overcome the drawback of a shortage of sign language data. Through experimental results, our proposed model outperforms a baseline BERT model by 6.22%.

Self-Supervised Learning for the Distinction between Computer-Graphics Images and Natural Images

With the increasing visual realism of computer-graphics (CG) images generated by advanced rendering engines, the distinction between CG images and natural images (NIs) has become an important research problem in the image forensics community. Previous research works mainly focused on the conventional supervised learning framework, which usually requires a good quantity of labeled data for training. To our knowledge, we study, for the first time in the literature, the utility of the self-supervised learning mechanism for the forensic classification of CG images and NIs. The idea is to make use of a large number of readily available unlabeled data, along with a self-supervised training procedure on a well-designed pretext task for which labels can be generated in an automatic and convenient way without human manual labeling effort. Differing from existing self-supervised methods, based on pretext tasks targeted at image understanding, or based on contrastive learning, we propose carrying out self-supervised training on a forensics-oriented pretext task of classifying authentic images and their modified versions after applying various manipulations. Experiments and comparisons showed the effectiveness of our method for solving the CG forensics problem under different evaluation scenarios. Our proposed method outperformed existing self-supervised methods in all experiments. It could sometimes achieve comparable, or better, performance. compared with a state-of-the-art fully supervised method under difficult evaluation scenarios with data scarcity and a challenging forensic problem. Our study demonstrates the utility and potential of the self-supervised learning mechanism for image forensics applications.

A Multi-strategy-based Pre-training Method for Cold-start Recommendation

The cold-start issue is a fundamental challenge in Recommender Systems. The recent self-supervised learning (SSL) on Graph Neural Networks (GNNs) model, PT-GNN, pre-trains the GNN model to reconstruct the cold-start embeddings and has shown great potential for cold-start recommendation. However, due to the over-smoothing problem, PT-GNN can only capture up to 3-order relation, which cannot provide much useful auxiliary information to depict the target cold-start user or item. Besides, the embedding reconstruction task only considers the intra-correlations within the subgraph of users and items, while ignoring the inter-correlations across different subgraphs. To solve the above challenges, we propose a multi-strategy-based pre-training method for cold-start recommendation (MPT), which extends PT-GNN from the perspective of model architecture and pretext tasks to improve the cold-start recommendation performance. 1 Specifically, in terms of the model architecture, in addition to the short-range dependencies of users and items captured by the GNN encoder, we introduce a Transformer encoder to capture long-range dependencies. In terms of the pretext task, in addition to considering the intra-correlations of users and items by the embedding reconstruction task, we add an embedding contrastive learning task to capture inter-correlations of users and items. We train the GNN and Transformer encoders on these pretext tasks under the meta-learning setting to simulate the real cold-start scenario, making the model able to be easily and rapidly adapted to new cold-start users and items. Experiments on three public recommendation datasets show the superiority of the proposed MPT model against the vanilla GNN models, the pre-training GNN model on user/item embedding inference, and the recommendation task.

Self-supervised visual representation learning on food images

Food image classification is the groundwork for image-based dietary assessment, which is the process of monitoring what kinds of food and how much energy is consumed using captured food or eating scene images. Existing deep learning based methods learn the visual representation for food classification based on human annotation of each food image. However, most food images captured from real life are obtained without labels, requiring human annotation to train deep learning based methods. This approach is not feasible for real world deployment due to high costs. To make use of the vast amount of unlabeled images, many existing works focus on unsupervised or self-supervised learning to learn the visual representation directly from unlabeled data. However, none of these existing works focuses on food images, which is more challenging than general objects due to its high inter-class similarity and intra-class variance. In this paper, we focus on two items: the comparison of existing models and the development of an effective self-supervised learning model for food image classification. Specifically, we first compare the performance of existing state-of-the-art self-supervised learning models, including SimSiam, SimCLR, SwAV, BYOL, MoCo, and Rotation Pretext Task on food images. The experiments are conducted on the Food-101 dataset, which contains 101 different classes of foods with 1,000 images in each class. Next, we analyze the unique features of each model and compare their performance on food images to identify the key factors in each model that can help improve the accuracy. Finally, we propose a new model for unsupervised visual representation learning on food images for the classification task.

Deep learning based on self-supervised pre-training: Application on sandstone content prediction

Deep learning has been widely used in various fields and showed promise in recent years. Therefore, deep learning is the future trend to realize seismic data’s intelligent and automatic interpretation. However, traditional deep learning only uses labeled data to train the model, and thus, does not utilize a large amount of unlabeled data. Self-supervised learning, widely used in Natural Language Processing (NLP) and computer vision, is an effective method of learning information from unlabeled data. Thus, a pretext task is designed with reference to Masked Autoencoders (MAE) to realize self-supervised pre-training of unlabeled seismic data. After pre-training, we fine-tune the model to the downstream task. Experiments show that the model can effectively extract information from unlabeled data through the pretext task, and the pre-trained model has better performance in downstream tasks.

Self-supervised learning for remote sensing scene classification under the few shot scenario

Scene classification is a crucial research problem in remote sensing (RS) that has attracted many researchers recently. It has many challenges due to multiple issues, such as: the complexity of remote sensing scenes, the classes overlapping (as a scene may contain objects that belong to foreign classes), and the difficulty of gaining sufficient labeled scenes. Deep learning (DL) solutions and in particular convolutional neural networks (CNN) are now state-of-the-art solution in RS scene classification; however, CNN models need huge amounts of annotated data, which can be costly and time-consuming. On the other hand, it is relatively easy to acquire large amounts of unlabeled images. Recently, Self-Supervised Learning (SSL) is proposed as a method that can learn from unlabeled images, potentially reducing the need for labeling. In this work, we propose a deep SSL method, called RS-FewShotSSL, for RS scene classification under the few shot scenario when we only have a few (less than 20) labeled scenes per class. Under this scenario, typical DL solutions that fine-tune CNN models, pre-trained on the ImageNet dataset, fail dramatically. In the SSL paradigm, a DL model is pre-trained from scratch during the pretext task using the large amounts of unlabeled scenes. Then, during the main or the so-called downstream task, the model is fine-tuned on the labeled scenes. Our proposed RS-FewShotSSL solution is composed of an online network and a target network both using the EfficientNet-B3 CNN model as a feature encoder backbone. During the pretext task, RS-FewShotSSL learns discriminative features from the unlabeled images using cross-view contrastive learning. Different views are generated from each image using geometric transformations and passed to the online and target networks. Then, the whole model is optimized by minimizing the cross-view distance between the online and target networks. To address the problem of limited computation resources available to us, our proposed method uses a novel DL architecture that can be trained using both high-resolution and low-resolution images. During the pretext task, RS-FewShotSSL is trained using low-resolution images, thereby, allowing for larger batch sizes which significantly boosts the performance of the proposed pipeline on the task of RS classification. In the downstream task, the target network is discarded, and the online network is fine-tuned using the few labeled shots or scenes. Here, we use smaller batches of both high-resolution and low-resolution images. This architecture allows RS-FewshotSSL to benefit from both large batch sizes and full image sizes, thereby learning from the large amounts of unlabeled data in an effective way. We tested RS-FewShotSSL on three RS public datasets, and it demonstrated a significant improvement compared to other state-of-the-art methods such as: SimCLR, MoCo, BYOL and IDSSL.

A bidirectional trajectory contrastive learning model for driving intention prediction

AbstractDriving intention prediction with trajectory data of surrounding vehicles is critical to advanced driver assistance system for improving the accuracy of decision-making. Previous works mostly focused on trajectory representation based on supervised manners. However, learning generalized and high-quality representations from unlabeled data remains a very challenging task. In this paper, we propose a self-supervised bidirectional trajectory contrastive learning (BTCL) model that learns generalized trajectory representation to improve the performance of the driving intention prediction task. Different trajectory data augmentation strategies and a cross-view trajectory prediction task are constructed jointly as pretext task of contrastive learning. The pretext task can maximize the similarity among different augmentations of the same sample while minimizing similarity among augmentations of different samples. It can not only learn the high-quality representation of trajectory without labeled information but also improve the adversarial attacks on BTCL. Moreover, considering the vehicle trajectory forward and backward follows the same social norms and driving behavior constraints. A bidirectional trajectory contrastive learning module is built to gain more positive samples that further increasing the prediction accuracy in downstream tasks and transfer ability of the model. Experimental results demonstrate that BTCL is competitive with the state-of-the-art, especially for adversarial attack and transfer learning tasks, on real-world HighD and NGSIM datasets.

Evolutionary Augmentation Policy Optimization for Self-Supervised Learning

Self-supervised Learning (SSL) is a machine learning algorithm for pretraining Deep Neural Networks (DNNs) without requiring manually labeled data. The central idea of this learning technique is based on an auxiliary stage aka pretext task in which labeled data are created automatically through data augmentation and exploited for pretraining the DNN. However, the effect of each pretext task is not well studied or compared in the literature. In this paper, we study the contribution of augmentation operators on the performance of self supervised learning algorithms in a constrained settings. We propose an evolutionary search method for optimization of data augmentation pipeline in pretext tasks and measure the impact of augmentation operators in several SOTA SSL algorithms. By encoding different combination of augmentation operators in chromosomes we seek the optimal augmentation policies through an evolutionary optimization mechanism. We further introduce methods for analyzing and explaining the performance of optimized SSL algorithms. Our results indicate that our proposed method can find solutions that outperform the accuracy of classification of SSL algorithms which confirms the influence of augmentation policy choice on the overall performance of SSL algorithms. We also compare optimal SSL solutions found by our evolutionary search mechanism and show the effect of batch size in the pretext task on two visual datasets.

Self-Supervised Modality-Aware Multiple Granularity Pre-Training for RGB-Infrared Person Re-Identification

RGB-Infrared person re-identification (RGB-IR ReID) aims to associate people across disjoint RGB and IR camera views. Currently, state-of-the-art performance of RGB-IR ReID is not as impressive as that of conventional ReID. Much of that is due to the notorious modality bias training issue brought by the single-modality ImageNet pre-training, which might yield RGB-biased representations that severely hinder the cross-modality image retrieval. This paper makes first attempt to tackle the task from a pre-training perspective. We propose a self-supervised pre-training solution, named Modality-Aware Multiple Granularity Learning (MMGL), which directly trains models from scratch only on multi-modal ReID datasets, but achieving competitive results against ImageNet pre-training, without using any external data or sophisticated tuning tricks. First, we develop a simple-but-effective ‘permutation recovery’ pretext task that globally maps shuffled RGB-IR images into a shared latent permutation space, providing modality-invariant global representations for downstream ReID tasks. Second, we present a part-aware cycle-contrastive (PCC) learning strategy that utilizes cross-modality cycle-consistency to maximize agreement between semantically similar RGB-IR image patches. This enables contrastive learning for the unpaired multi-modal scenarios, further improving the discriminability of local features without laborious instance augmentation. Based on these designs, MMGL effectively alleviates the modality bias training problem. Extensive experiments demonstrate that it learns better representations (+8.03% Rank-1 accuracy) with faster training speed (converge only in few hours) and higher data efficiency (<5% data size) than ImageNet pre-training. The results also suggest it generalizes well to various existing models, losses and has promising transferability across datasets. The code will be released at https://github.com/hansonchen1996/MMGL.

Nonintrusive Load Monitoring Based on Self-Supervised Learning

Deep learning models for non-intrusive load monitoring (NILM) tend to require a large amount of labeled data for training. However, it is difficult to generalize the trained models to unseen sites due to different load characteristics and operating patterns of appliances between data sets. For addressing such problems, self-supervised learning (SSL) is proposed in this paper, where labeled appliance-level data from the target data set or house is not required. Initially, only the aggregate power readings from target data set are required to pre-train a general network via a self-supervised pretext task to map aggregate power sequences to derived representatives. Then, supervised downstream tasks are carried out for each appliance category to fine-tune the pre-trained network, where the features learned in the pretext task are transferred. Utilizing labeled source data sets enables the downstream tasks to learn how each load is disaggregated, by mapping the aggregate to labels. Finally, the fine-tuned network is applied to load disaggregation for the target sites. For validation, multiple experimental cases are designed based on three publicly accessible REDD, UK-DALE, and REFIT data sets. Besides, state-of-the-art neural networks are employed to perform NILM task in the experiments. Based on the NILM results in various cases, SSL generally outperforms zero-shot learning in improving load disaggregation performance without any sub-metering data from the target data sets.

Self-Supervised Feature Enhancement: Applying Internal Pretext Task to Supervised Learning

Traditional self-supervised learning requires CNNs using external pretext tasks (i.e., image- or video-based tasks) to encode high-level semantic visual representations. In this paper, we show that feature transformations within CNNs can also be regarded as supervisory signals to construct the self-supervised task, called \emph{internal pretext task}. And such a task can be applied for the enhancement of supervised learning. Specifically, we first transform the internal feature maps by discarding different channels, and then define an additional internal pretext task to identify the discarded channels. CNNs are trained to predict the joint labels generated by the combination of self-supervised labels and original labels. By doing so, we let CNNs know which channels are missing while classifying in the hope to mine richer feature information. Extensive experiments show that our approach is effective on various models and datasets. And it's worth noting that we only incur negligible computational overhead. Furthermore, our approach can also be compatible with other methods to get better results.

Object-Centric Masked Image Modeling-Based Self-Supervised Pretraining for Remote Sensing Object Detection

Masked image modeling (MIM) has been proved to be an optimal pretext task for self-supervised pretraining (SSP), which can facilitate the model to capture an effective task-agnostic representation at the pretraining step and then advance the fine-tuning performance of various downstream tasks. However, under the high randomly masked ratio of MIM, the scene-level MIM-based SSP is hard to capture the small-scale objects or local details from complex remote sensing scenes. Then, when the pretrained models capturing more scene-level information are directly applied for object-level fine-tuning step, there is an obvious representation learning misalignment between model pretraining and fine-tuning steps. Therefore, in this article, a novel object-centric masked image modeling (OCMIM) strategy is proposed to make the model better capture the object-level information at the pretraining step and then further advance the object detection fine-tuning step. First, to better learn the object-level representation involving full scales and multicategories at MIM-based SSP, a novel object-centric data generator is proposed to automatically setup targeted pretraining data according to objects themselves, which can provide the specific data condition for object detection model pretraining. Second, an attention-guided mask generator is designed to generate a guided mask for MIM pretext task, which can lead the model to learn more discriminative representation of highly attended object regions than by using the randomly masking strategy. Finally, several experiments are conducted on six remote sensing object detection benchmarks, and results proved that the proposed OCMIM-based SSP strategy is a better pretraining way for remote sensing object detection than normally used methods.

MTHM: Self-Supervised Multitask Anomaly Detection With Hard Example Mining

It is still challenging to detect and locate anomalies by models trained only with normal samples. Methods using image reconstruction as a pretext task can provide precise localization but suffer from harnessing the reconstruction capability on unseen anomalies. This paper proposes a new framework of Multi-Task and Hard example Mining (MTHM) for anomaly detection and localization. The self-supervised multi-task setting creatively takes advantage of the competition among different tasks to learn more compact and efficient representations for detection tasks. Moreover, introducing other semantic tasks allows the shared encoder to learn beyond the pixel-to-pixel mapping of only a single image reconstruction task. Subsequent analysis experiments demonstrate that the proposed method can achieve a more suppressive reconstruction capability for anomalies. During the test process, the outputs of the other tasks can also provide valuable information for anomaly detection and localization. Furthermore, in combination with a novel hard example mining strategy, the byproducts of the image reconstruction task are inexpensively exploited as hard-to-detect samples for enhancing models’ detection ability. And as the model capability increases during training, the detection difficulty of these samples is able to increase adaptively. Our experiments show that hard samples generated in the later training stages can better approximate the real data distribution. With the help of the multi-task framework and hard example mining strategy, our method surpasses many state-of-the-art methods.

Masked Spectral Bands Modeling With Shifted Windows: An Excellent Self-Supervised Learner for Classification of Medical Hyperspectral Images

Hyperspectral imaging has become a popular imaging technique in the medical field, and the development of algorithms for computer-aided diagnosis (CAD) is urgently required. Traditional deep learning techniques require a lot of annotated data, which is a burden on doctors. Self-supervised learning (SSL) is a solution for extracting feature representations from unlabeled data. However, traditional CNN-based SSL algorithms cannot explore relations between neighboring and long-range spectral bands, which limits classification performance. In this letter, the proposed solution is a novel SSL method using a transformer-based technique called masked spectral bands modeling with shifted windows (MSBMSW). This method predicts masked spectral bands as the pretext task and uses a self-attention mechanism with shifted windows to capture the divergence of neighboring spectral bands and enhance information exchange between long-range spectral bands. Experimental results demonstrate that MSBMSW achieves better classification results than many state-of-the-art methods and has potential clinical value for CAD of MHSIs.