Search

Air access networks have been recognized as a significant driver of various Internet of Things (IoT) services and applications. In particular, the aerial computing network infrastructure centered on the Internet of Drones has set off a new revolution in automatic image recognition. This emerging technology relies on sharing ground truth labeled data between Unmanned Aerial Vehicle (UAV) swarms to train a high-quality automatic image recognition model. However, such an approach will bring data privacy and data availability challenges. To address these issues, we first present a Semi-supervised Federated Learning (SSFL) framework for privacy-preserving UAV image recognition. Specifically, we propose model parameters mixing strategy to improve the naive combination of FL and semi-supervised learning methods under two realistic scenarios (labels-at-client and labels-at-server), which is referred to as Federated Mixing (FedMix). Furthermore, there are significant differences in the number, features, and distribution of local data collected by UAVs using different camera modules in different environments, i.e., statistical heterogeneity. To alleviate the statistical heterogeneity problem, we propose an aggregation rule based on the frequency of the client's participation in training, namely the FedFreq aggregation rule, which can adjust the weight of the corresponding local model according to its frequency. Numerical results demonstrate that the performance of our proposed method is significantly better than those of the current baseline and is robust to different non-IID levels of client data., Comment: arXiv admin note: text overlap with arXiv:2110.13388

The scarcity of labels is common and great challenge in traditional supervised learning. Semi-supervised learning (SSL) leverages unlabeled samples to alleviate the absence of label information. Similar with annotation, label proportion is another type of prior information and plays a significant role in classification tasks. Compared with the acquisition of labels, label proportion can be obtained more easily. For example, only a small number of patients have been diagnosed with or not with cancers in hospital database, while the proportion with cancer can be generally estimated by historical records. How to incorporate such prior information of label proportion is crucial but rarely studied in literature. Traditional SSL methods often ignore this prior information and will lead to performance degradation inevitably. To solve this problem, we propose a novel SSL with Label Proportion (SSLLP). Our approach encourages to preserve label consistency and label proportion by imposing the cardinality bound constraints. Our formulated problem equals to a mixed-integer constrained submodular minimization and it is difficult to be solved directly. Therefore, we transformed the original problem into a convex one by Lov $\acute{\text{a}}$ sz extension and designed an efficient solving algorithm. Extensive experimental results present the improved performance of our method over several state-of-the-art methods.

Deep learning methodologies have been employed in several different fields, with an outstanding success in image recognition applications, such as material quality control, medical imaging, autonomous driving, etc. Deep learning models rely on the abundance of labelled observations to train a prospective model. These models are composed of millions of parameters to estimate, increasing the need of more training observations. Frequently it is expensive to gather labelled observations of data, making the usage of deep learning models not ideal, as the model might over-fit data. In a semi-supervised setting, unlabelled data is used to improve the levels of accuracy and generalization of a model with small labelled datasets. Nevertheless, in many situations different unlabelled data sources might be available. This raises the risk of a significant distribution mismatch between the labelled and unlabelled datasets. Such phenomena can cause a considerable performance hit to typical semi-supervised deep learning frameworks, which often assume that both labelled and unlabelled datasets are drawn from similar distributions. Therefore, in this paper we study the latest approaches for semi-supervised deep learning for image recognition. Emphasis is made in semi-supervised deep learning models designed to deal with a distribution mismatch between the labelled and unlabelled datasets. We address open challenges with the aim to encourage the community to tackle them, and overcome the high data demand of traditional deep learning pipelines under real-world usage settings., Comment: Submission to IEEE Transactions on AI

Federated Learning is a distributed machine learning paradigm dealing with decentralized and personal datasets. Since data reside on devices such as smartphones and virtual assistants, labeling is entrusted to the clients or labels are extracted in an automated way. Specifically, in the case of audio data, acquiring semantic annotations can be prohibitively expensive and time-consuming. As a result, an abundance of audio data remains unlabeled and unexploited on users’ devices. Most existing federated learning approaches focus on supervised learning without harnessing the unlabeled data. In this work, we study the problem of semi-supervised learning of audio models via self-training in conjunction with federated learning. We propose FedSTAR to exploit large-scale on-device unlabeled data to improve the generalization of audio recognition models. We further demonstrate that self-supervised pre-trained models can accelerate the training of on-device models, significantly improving convergence within fewer training rounds. We conduct experiments on diverse public audio classification datasets and investigate the performance of our models under varying percentages of labeled and unlabeled data. Notably, we show that with as little as 3% labeled data available, FedSTAR on average can improve the recognition rate by 13.28% compared to the fully supervised federated model.

Continual Learning (CL) investigates how to train Deep Networks on a stream of tasks without incurring forgetting. CL settings proposed in literature assume that every incoming example is paired with ground-truth annotations. However, this clashes with many real-world applications: gathering labeled data, which is in itself tedious and expensive, becomes infeasible when data flow as a stream. This work explores Continual Semi-Supervised Learning (CSSL): here, only a small fraction of labeled input examples are shown to the learner. We assess how current CL methods (e.g.: EWC, LwF, iCaRL, ER, GDumb, DER) perform in this novel and challenging scenario, where overfitting entangles forgetting. Subsequently, we design a novel CSSL method that exploits metric learning and consistency regularization to leverage unlabeled examples while learning. We show that our proposal exhibits higher resilience to diminishing supervision and, even more surprisingly, relying only on 25% supervision suffices to outperform SOTA methods trained under full supervision., Comment: 7 pages, 2 figures, to appear in Pattern Recognition Letters, Volume 162, October 2022, Pages 9-14

Due to the shortage of COVID-19 viral testing kits, radiology is used to complement the screening process. Deep learning methods are promising in automatically detecting COVID-19 disease in chest x-ray images. Most of these works first train a Convolutional Neural Network (CNN) on an existing large-scale chest x-ray image dataset and then fine-tune the model on the newly collected COVID-19 chest x-ray dataset, often at a much smaller scale. However, simple fine-tuning may lead to poor performance due to two issues, firstly the large domain shift present in chest x-ray datasets and secondly the relatively small scale of the COVID-19 chest x-ray dataset. In an attempt to address these issues, we formulate the problem of COVID-19 chest x-ray image classification in a semi-supervised open set domain adaptation setting and propose a novel domain adaptation method, Semi-supervised Open set Domain Adversarial network (SODA). SODA is designed to align the data distributions across different domains in the general domain space and also in the common subspace of source and target data. In our experiments, SODA achieves a leading classification performance compared with recent state-of-the-art models in separating COVID-19 with common pneumonia. We also present results showing that SODA produces better pathology localizations.

MicroRNAs (miRNAs) are a class of non-coding RNAs that play critical role in many biological processes, such as cell growth, development, differentiation and aging. Increasing studies have revealed that miRNAs are closely involved in many human diseases. Therefore, the prediction of miRNA-disease associations is of great significance to the study of the pathogenesis, diagnosis and intervention of human disease. However, biological experimentally methods are usually expensive in time and money, while computational methods can provide an efficient way to infer the underlying disease-related miRNAs. In this study, we propose a novel method to predict potential miRNA-disease associations, called SVAEMDA. Our method mainly consider the miRNA-disease association prediction as semi-supervised learning problem. SVAEMDA integrates disease semantic similarity, miRNA functional similarity and respective Gaussian interaction profile (GIP) similarities. The integrated similarities are used to learn the representations of diseases and miRNAs. SVAEMDA trains a variational autoencoder based predictor by using known miRNA-disease associations, with the form of concatenated dense vectors. Reconstruction probability of the predictor is used to measure the correlation of the miRNA-disease pairs. Experimental results show that SVAEMDA outperforms other stat-of-the-art methods. AUC values of SVAEMDA of global leave-one-out cross validation (LOOCV) and 5-fold cross validation (5-fold CV) are 0.9464 and 0.9428 respectively. In addition, case studies of three common human diseases indicate that SVAEMDA obtains 100 percent of the top 50 predicted candidates in the benchmark databases. Therefore, SVAEMDA can efficiently and accurately predict the potential associations between diseases and miRNAs.

As residential photovoltaic (PV) system installations continue to increase rapidly, utilities need to identify the locations of these new components to manage the unconventional two-way power flow and maintain sustainable management of distribution grids. But historical records are unreliable and constant re-assessment of active residential PV locations is resource intensive. To resolve these issues, we propose to model the solar detection problem in a machine learning set up based on labeled data, e.g., supervised learning. However, the challenge for most utilities is limited labels or labels on only one type of users. Therefore, we design new semi-supervised learning and one-class classification methods based on autoencoders, which greatly improve the nonlinear data representation of human behavior and solar behavior. The proposed methods have been tested and validated not only on synthetic data based on a publicly available dataset, but also on real-world data from utility partners. The numerical results show robust detection accuracy, laying down the foundation for managing distributed energy resources in distribution grids.

To characterize a community-scale urban functional area using geo-tagged data and available land-use information, several supervised and semi-supervised classification models are presented and evaluated in Hong Kong for comparing their uncertainty, robustness and sensitivity. The following results are noted: (i) As the training set size grows, models’ accuracies are improved, particularly for multi-layer perceptron (MLP) or random forest (RF). The graph convolutional network (GCN) (MLP or RF) model reveals top accuracy when the proportion of training samples is less (greater) than 10% of the total number of functional areas; (ii) With a large amount of training samples, MLP shows the highest prediction accuracy and good performances in cross-validation, but less stability on same training sets; (iii) With a small amount of training samples, GCN provides viable results, by incorporating the auxiliary information provided by the proposed semantic linkages, which is meaningful in real-world predictions; (iv) When the training samples are less than 10%, one should be cautious using MLP to test the optimal epoch for obtaining the best accuracy, due to its model overfitting problem. The above insights could support efficient and scalable urban functional area mapping, even with insufficient land-use information (e.g., covering only ~20% of Beijing in the case study).

Producing densely annotated data is a difficult and tedious task for medical imaging applications. To address this problem, we propose a novel approach to generate supervision for semi-supervised semantic segmentation. We argue that visually similar regions between labeled and unlabeled images likely contain the same semantics and therefore should share their label. Following this thought, we use a small number of labeled images as reference material and match pixels in an unlabeled image to the semantics of the best fitting pixel in a reference set. This way, we avoid pitfalls such as confirmation bias, common in purely prediction-based pseudo-labeling. Since our method does not require any architectural changes or accompanying networks, one can easily insert it into existing frameworks. We achieve the same performance as a standard fully supervised model on X-ray anatomy segmentation, albeit 95% fewer labeled images. Aside from an in-depth analysis of different aspects of our proposed method, we further demonstrate the effectiveness of our reference-guided learning paradigm by comparing our approach against existing methods for retinal fluid segmentation with competitive performance as we improve upon recent work by up to 15% mean IoU., 36th AAAI Conference on Artificial Intelligence 2022

The article deals with the problem of noise effect on semi-supervised learning. The goal of this article is to analyze the impact of noise on the accuracy of binary classification models created using three semi-supervised learning algorithms, namely Simple Recycled Selection, Incrementally Reinforced Selection, and Hybrid Algorithm, using Support Vector Machines to build a base classifier. Different algorithms to compute similarity matrices, namely Radial Bias Function, Cosine Similarity, and K-Nearest Neighbours were analyzed to understand their effect on model accuracy. For benchmarking purposes, datasets from the UCI repository were used. To test the noise effect, different amounts of artificially generated randomly-labeled samples were introduced into the dataset using three strategies (labeled, unlabeled, and mixed) and compared to the baseline classifier trained with the original dataset and the classifier trained on the reduced-size original dataset. The results show that the introduction of random noise into the labeled samples decreases classifier accuracy, while a moderate amount of noise in unmarked samples can have a positive effect on classifier accuracy. В статье рассматривается проблема влияния шума на точность в задачах полууправляемого обучения. Целью этой статьи является анализ влияния шума на точность моделей, созданных с помощью трех полууправляемых алгоритмов обучения, а именно: Simply Recycled Selection (SRS), Incrementally Reinforced Selection (IRS) и Hybrid Algorithm (HYB). В качестве базового классификатора использован Support Vector Machine (SVM). Проанализированы различные алгоритмы для вычисления матриц подобия, а именно Radial Bias Function, Cosine Simmilarity и K-Nearest Neighbours. Для целей сравнительного анализа будут использоваться наборы данных из хранилища UCI. Чтобы проверить влияние шумов, различные количества искусственно сгенерированных случайно размеченных образцов были введены в набор данных с использованием трех стратегий (размеченные, неразмеченные, смешанные) и сравнены с базовым классификатором, обученным исходным набором данных, и классификатором, обученным на исходном наборе данных уменьшенного размера. Результаты показывают, что ввод случайного шума в маркированные образцы ухудшает точность модели, а ввод случайного шума в немаркированные данные может наоборот повысить точность модели. У статті розглядається проблема впливу шуму на точність у задачах напівккерованого навчання. Метою цієї статті є аналіз впливу шуму на точність моделей бінарної класифікації, створених за допомогою трьох напівкерованих алгоритмів навчання, а саме: Simply Recycled Selection (SRS), Incrementally Reinforced Selection (IRS) і Hybrid Algorithm (HYB). У якості базового класифікатора використано Support Vector Machine (SVM). Ми проаналізуємо різні алгоритми для обчислення матриць подібності, а саме Radial Bias Function, Cosine Simmilarity і K-Nearest Neighbours. Для цілей порівняльного аналізу використовуватимуться набори даних зі сховища UCI. Щоб перевірити вплив шуму, різна кількість штучно згенерованих випадково позначених зразків було введено в набір даних з використанням трьох стратегій (маркована, не маркована та змішана) і порівняно з базовим класифікатором, навченим з вихідним набором даних, і класифікатором, навченим на вихідному наборі даних зменшеного розміру. Результати показують, що введення випадкового шуму в марковані зразки погіршує точність моделі, а введення випадкового шуму в немарковані дані може навпаки підвищити точність моделі. Ключові слова: зашумлені данні; машинне навчання; напівкероване навчання; опорні векторні машини.

Knowledge graphs have received intensive research interests. When the labels of most nodes or datapoints are missing, anchor graph and hierarchical anchor graph models can be employed. With an anchor graph or hierarchical anchor graph, we only need to optimize the labels of the coarsest anchors, and the labels of datapoints can be inferred from these anchors in a coarse-to-fine manner. The complexity of optimization is therefore reduced to a cubic cost with respect to the number of the coarsest anchors. However, to obtain a high accuracy when a data distribution is complex, the scale of this anchor set still needs to be large, which thus inevitably incurs an expensive computational burden. As such, a challenge in scaling up these models is how to efficiently estimate the labels of these anchors while keeping classification performance. To address this problem, we propose a novel approach that adds an anchor label predictor in the conventional anchor graph and hierarchical anchor graph models. In the proposed approach, the labels of the coarsest anchors are not directly optimized, and instead, we learn a label predictor which estimates the labels of these anchors with their spectral representations. The predictor is optimized with a regularization on all datapoints based on a hierarchical anchor graph, and we show that its solution only involves the inversion of a small-size matrix. Built upon the anchor hierarchy, we design a sparse intra-layer adjacency matrix over these anchors, which can simultaneously accelerate spectral embedding and enhance effectiveness. Our approach is named Faster Learning6 on Anchor Graph (FLAG) as it improves conventional anchor-graph-based methods in terms of efficiency. Experiments on a variety of publicly available datasets with sizes varying fromthousands to millions of Q1 samples demonstrate the effectiveness of our approach.

Semi-supervised learning aims to learn prediction models from both labeled and unlabeled samples. There has been extensive research in this area. Among existing work, generative mixture models with Expectation-Maximization (EM) is a popular method due to clear statistical properties. However, existing literature on EM-based semi-supervised learning largely focuses on unstructured prediction, assuming that samples are independent and identically distributed. Studies on EM-based semi-supervised approach in structured prediction is limited. This paper aims to fill the gap through a comparative study between unstructured and structured methods in EM-based semi-supervised learning. Specifically, we compare their theoretical properties and find that both methods can be considered as a generalization of self-training with soft class assignment of unlabeled samples, but the structured method additionally considers structural constraint in soft class assignment. We conducted a case study on real-world flood mapping datasets to compare the two methods. Results show that structured EM is more robust to class confusion caused by noise and obstacles in features in the context of the flood mapping application.

In this paper, we propose a new semi-supervised graph construction method, which is capable of adaptively learning the similarity relationship between data samples by fully exploiting the potential of pairwise constraints, a kind of weakly supervisory information. Specifically, to adaptively learn the similarity relationship, we linearly approximate each sample with others under the regularization of the low-rankness of the matrix formed by the approximation coefficient vectors of all the samples. In the meanwhile, by taking advantage of the underlying local geometric structure of data samples that is empirically obtained, we enhance the dissimilarity information of the available pairwise constraints via propagation. We seamlessly combine the two adversarial learning processes to achieve mutual guidance. We cast our method as a constrained optimization problem and provide an efficient alternating iterative algorithm to solve it. Experimental results on five commonly-used benchmark datasets demonstrate that our method produces much higher classification accuracy than state-of-the-art methods, while running faster.

Hyperspectral image (HSI) classification is gaining a lot of momentum in present time because of high inherent spectral information within the images. However, these images suffer from the problem of curse of dimensionality and usually require a large number samples for tasks such as classification, especially in supervised setting. Recently, to effectively train the deep learning models with minimal labelled samples, the unlabeled samples are also being leveraged in self-supervised and semi-supervised setting. In this work, we leverage the idea of semi-supervised learning to assist the discriminative self-supervised pretraining of the models. The proposed method takes different augmented views of the unlabeled samples as input and assigns them the same pseudo-label corresponding to the labelled sample from the downstream task. We train our model on two HSI datasets, namely Houston dataset (from data fusion contest, 2013) and Pavia university dataset, and show that the proposed approach performs better than self-supervised approach and supervised training., Comment: The paper was submitted in IGARSS, 2023 conference and is not accepted to appear in the proceedings. The page requirement is 4 pages, including references

In practical scenarios, a sample may have multiple labels that reveal its classes instead of a single label, which is widely known as multi-label classification (MLC). However, some practical situations may lack reliable labels due to the high cost, time-consuming and professional labelling process. Although Semi-supervised classification may become a potential solution, most of the outstanding existing methods are customized for the single-label situation and ignore multi-label situations. Consistency regularization has performed great success in Weakly/Semi-supervised Single-label classification (SS-SLC), but few efforts have been devoted to semi-supervised Multi-label classification (SS-MLC). A simple solution for introducing consistency regularization to SS-MLC is to regularize predictions of models to be consistent with different augmentation of the same image. Nonetheless, the solution lacks attention to label relations which are crucial components of Multi-label classification. In the thesis, I go beyond the consistency regularization in SS-SLC and propose Conditional Consistency Regularization (CCR) that is designed for SS-MLC. To be specific, we make potential labels (grand-truth label for labeled samples, pseudo-label for unlabeled samples) conditioned on different label states (i.e., positive, negative, or unknown for each class). By regularizing the two predictions to be invariant, the model can learn label relations implicitly between two different label states, which can boost classification performance. The comprehensive experiments that are conducted on different datasets show that the proposed method can surpass state-of-art SS-MLC and MLC methods by a large gap.

International audience; Co-clustering aims at simultaneously partitioning both dimensions of a data matrix. It has demonstrated better performances than one-sided clustering for high-dimensional data. The Latent Block Model (LBM) is a probabilistic model for co-clustering based on mixture models that has proven useful for a broad class of data. In this paper, we propose to leverage prior knowledge in the form of pairwise semi-supervision in both row and column space to improve the clustering performances of the algorithms derived from this model. We present a general probabilistic framework for incorporating Must Link and Cannot Link relationships in the LBM based on Hidden Markov Random Fields (HMRF). We instantiate this framework on a model for count data and present two inference algorithms based on Variational and Classification EM. Extensive experiments on simulated data and on real-world attributed networks confirm the interest of our approach and demonstrate the effectiveness of our algorithms.

Traditional active learning algorithms have several limitations: 1) they cannot obtain satisfactory results on high dimensional datasets, especially for multivariate time series (MTS) data; 2) traditional crowd-based labeling approaches do not consider the swarm intelligence of crowds, which cannot guarantee the confidence of labeling results; and 3) up to now, few works have addressed the issue of tradeoff between the labeling accuracy and the labeling cost with crowdsourcing labelers. There is also a lack of research on crowd-based active learning on MTS data. To efficiently address the above issues, we propose a new framework of active semi-supervised learning. First, two criteria are advanced to measure the importance of an unlabeled sample from different perspectives. A dynamic time wrapping (DTW)-based similarity matrix is used to represent the MTS data. Next, to confidently label the most valuable samples with minimum cost, we advance a cost-effective crowd selection model and an adaptive labeler selection approach (ALS) to select the most suitable labelers, which could minimize the total cost of labeling and achieve better classification performance. Experiments on ten datasets show the effectiveness of our proposed methods.

Early classification of ongoing alarm floods in industrial monitoring systems is crucial to provide a safe and efficient operation. In this article, a data-driven approach is proposed to address the early classification problem with unlabeled historical data. To prioritize earlier activated alarms and take advantage of the triggering time information of alarms, a vector representation called exponentially attenuated component (EAC) is used to represent alarm floods. This makes alarm sequences fit for different powerful machine learning algorithms. A method based on the time information of unlabeled historical alarm floods is formulated to determine the attenuation coefficient for EAC representation. With the Gaussian mixture model, an efficient semi-supervised approach is proposed to provide an early classification of alarm floods using unlabeled historical data. The efficiency of the proposed method is validated by the Tennessee Eastman process benchmark and a real industrial dataset.

GPS trajectories serve as a significant data source for travel mode identification along with the development of various GPS-enabled smart devices. However, such data directly integrate user private information, thus hindering users from sharing data with third parties. On the other hand, existing identification methods heavily depend on the respective manual travel mode annotations, whose production is economically inefficient and error-prone. In this paper, we propose a Semi-supervised Federated Learning (SSFL) framework that can accurately identify travel modes without using users' raw trajectories data or relying on notable data labels. Specifically, we propose a new identification model named convolutional neural network-gated recurrent unit model in SSFL to accurately infer travel modes from GPS trajectories. Second, we design a pseudo-labeling method for the clients to set pseudo-labels on their local unlabeled dataset by using a small public dataset at the server. Furthermore, we adopt a grouping-based aggregation scheme and a data flipping augmentation scheme, which can boost the convergence and performance of the proposed framework. Comprehensive evaluations on a real-world dataset show that SSFL outperforms centralized semi-supervised baselines and is robust to the non-independent and identically distributed data commonly seen in practice.

Semi-supervised nonnegative matrix factorization (NMF) has received considerable attention in machine learning and data mining. A new semi-supervised NMF method, called dual semi-supervised convex nonnegative matrix factorization (DCNMF), is proposed in this paper for fully using the limited label information. Specifically, DCNMF simultaneously incorporates the pointwise and pairwise constraints of labeled samples as dual supervisory information into convex NMF, which results in a better low-dimensional data representation. Moreover, DCNMF imposes the nonnegative constraint only on the coefficient matrix but not on the base matrix. Consequently, DCNMF can process mixed-sign data, and hence enlarge the range of applications. We derive an efficient alternating iterative algorithm for DCNMF to solve the optimization, and analyze the proposed DCNMF method in terms of the convergence and computational complexity. We also discuss the relationships between DCNMF and several typical NMF based methods. Experimental results illustrate that DCNMF outperforms the related state-of-the-art NMF methods on nonnegative and mixed-sign datasets for clustering applications. This work is supported in part by the National Nature Science Foundation of China (No. U1701266, 91648208, 61976175), and Guangdong Intellectual Property Big Data Key Laboratory (No. 2018B030322016).

Semi-supervised learning provides great significance in left atrium (LA) segmentation model learning with insufficient labelled data. Generalising semi supervised learning to cross-domain data is of high importance to further improve model robustness. However, the widely existing distribution difference and sample mismatch between different data domains hinder the generalisation of semi-supervised learning. In this study, we alleviate these problems by proposing an Adaptive Hier10 archical Dual Consistency (AHDC) for the semi-supervised LA segmentation on cross-domain data. The AHDC mainly consists of a Bidirectional Adversarial Inference module (BAI) and a Hierarchical Dual Consistency learning module (HDC). The BAI overcomes the difference of distributions and the sample mismatch between two different domains. It mainly learns two mapping networks adversarially to obtain two matched domains through mutual adaptation. The HDC investigates a hierarchical dual learning paradigm for cross-domain semi-supervised segmentation based on the obtained matched domains. It mainly builds two dual modelling networks for mining the complementary information in both intra-domain and inter-domain. For the intra domain learning, a consistency constraint is applied to the dual-modelling targets to exploit the complementary modelling information. For the inter-domain learning, a consistency constraint is applied to the LAs modelled by two dual modelling networks to exploit the complementary knowl28 edge among different data domains. We demonstrated the performance of our proposed AHDC on four 3D late gadolinium enhancement cardiac MR (LGE-CMR) datasets from different centres and a 3D CT dataset. Compared to other state-of-the-art methods, our proposed AHDC achieved higher segmentation accuracy, which indicated its capability in the cross-domain semi-supervised LA segmentation.

Medical instrument segmentation in 3D ultrasound is essential for image-guided intervention. However, to train a successful deep neural network for instrument segmentation, a large number of labeled images are required, which is expensive and time-consuming to obtain. In this article, we propose a semi-supervised learning (SSL) framework for instrument segmentation in 3D US, which requires much less annotation effort than the existing methods. To achieve the SSL learning, a Dual-UNet is proposed to segment the instrument. The Dual-UNet leverages unlabeled data using a novel hybrid loss function, consisting of uncertainty and contextual constraints. Specifically, the uncertainty constraints leverage the uncertainty estimation of the predictions of the UNet, and therefore improve the unlabeled information for SSL training. In addition, contextual constraints exploit the contextual information of the training images, which are used as the complementary information for voxel-wise uncertainty estimation. Extensive experiments on multiple ex-vivo and in-vivo datasets show that our proposed method achieves Dice score of about 68.6%-69.1% and the inference time of about 1 sec. per volume. These results are better than the state-of-the-art SSL methods and the inference time is comparable to the supervised approaches., Comment: Accepted by IEEE JBHI

In a dynamic stream there is an assumption that the underlying process generating the stream is non-stationary and that concepts within the stream will drift and change as the stream progresses. Concepts learned by a classification model are prone to change and non-adaptive models are likely to deteriorate and become ineffective over time. The challenge of recognising and reacting to change in a stream is compounded by the scarcity of labels problem. This refers to the very realistic situation in which the true class label of an incoming point is not immediately available (or might never be available) or in situations where manually annotating data points are prohibitively expensive. In a high-velocity stream, it is perhaps impossible to manually label every incoming point and pursue a fully supervised approach. In this article, we formally describe the types of change, which can occur in a data-stream and then catalogue the methods for dealing with change when there is limited access to labels. We present an overview of the most influential ideas in the field along with recent advancements and we highlight trends, research gaps, and future research directions.

Vehicle re-identification (Re-ID) methods often fail to achieve robust performance due to insufficient training data and domain diversities. Although state-of-the-art methods apply image-to-image translation or web data to achieve data augmentation, the construct of new datasets will not only introduce noise, but also undergo a mismatch issue with the source domain. Moreover, the label noise of cross-domain data in existing label distribution technologies cannot be alleviated. In this paper, a multi-domain joint learning with inter-domain adaptation label smoothing regularization (IALSR) is proposed using a semi-supervised learning framework. The overall framework consists of two parts. In one part, a multi-domain joint network (MJNet) is proposed to learn multiple vehicle attributes simultaneously. The output of the training model is employed to group several inter-domain subsets, which are regarded as different domains. To adapt to domain diversities, style transfer models are learned for each pair of subsets to generate free and rich data as a novel data augmentation approach. In the other part, IALSR, which preserves self-similarity and domain-transitivity, is designed to smooth the noise of style-transferred data. Upon our basis, we further introduce the web data to verify the superiority of the IALSR. The results of extensive experimental on two large-scale vehicle Re-ID datasets demonstrate that the proposed approach is superior to other state-of-the-art ones.

A drug-drug interaction (DDI) is defined as an association between two drugs where the pharmacological effects of a drug are influenced by another drug. Positive DDIs can usually improve the therapeutic effects of patients, but negative DDIs cause the major cause of adverse drug reactions and even result in the drug withdrawal from the market and the patient death. Therefore, identifying DDIs has become a key component of the drug development and disease treatment. In this study, we propose a novel method to predict DDIs based on the integrated similarity and semi-supervised learning (DDI-IS-SL). DDI-IS-SL integrates the drug chemical, biological and phenotype data to calculate the feature similarity of drugs with the cosine similarity method. The Gaussian Interaction Profile kernel similarity of drugs is also calculated based on known DDIs. A semi-supervised learning method (the Regularized Least Squares classifier) is used to calculate the interaction possibility scores of drug-drug pairs. In terms of the 5-fold cross validation, 10-fold cross validation and de novo drug validation, DDI-IS-SL can achieve the better prediction performance than other comparative methods. In addition, the average computation time of DDI-IS-SL is shorter than that of other comparative methods. Finally, case studies further demonstrate the performance of DDI-IS-SL in practical applications.

Class-Incremental Learning (CIL) aims at incrementally learning novel classes without forgetting old ones. This capability becomes more challenging when novel tasks contain one or a few labeled training samples, which leads to a more practical learning scenario, i.e., Few-Shot Class-Incremental Learning (FSCIL). The dilemma on FSCIL lies in serious overfitting and exacerbated catastrophic forgetting caused by the limited training data from novel classes. In this paper, excited by the easy accessibility of unlabeled data, we conduct a pioneering work and focus on a Semi-Supervised Few-Shot Class-Incremental Learning (Semi-FSCIL) problem, which requires the model incrementally to learn new classes from extremely limited labeled samples and a large number of unlabeled samples. To address this problem, a simple but efficient framework is first constructed based on the knowledge distillation technique to alleviate catastrophic forgetting. To efficiently mitigate the overfitting problem on novel categories with unlabeled data, uncertainty-guided semi-supervised learning is incorporated into this framework to select unlabeled samples into incremental learning sessions considering the model uncertainty. This process provides extra reliable supervision for the distillation process and contributes to better formulating the class means. Our extensive experiments on CIFAR100, miniImageNet and CUB200 datasets demonstrate the promising performance of our proposed method, and define baselines in this new research direction.

Multi-view semi-supervised learning (SSL) has attracted great attention due to its effectiveness in information utilization of multiple views and labeled and unlabeled data to solve practical problems. However, most existing methods exhibit high computational complexity. Effective integration of the information on different views to achieve enhanced performance remains a challenging task. In this study, we combine an anchor-based approach with multi-view semi-supervised learning to address these problems. A novel multi-view SSL method called fast multi-view SSL (FMSSL) based on learned graph is proposed. Starting from the affinity graphs constructed by using an anchor-based strategy, FMSSL learns an optimal multi-view consensus graph by using feature and label information. The learned graph can jointly consider the relation of multiple views to approximate the manifold structure. The learned graph is then introduced into the SSL model as the weight matrix of a bipartite graph to simultaneously perform separate classification on the original samples and anchors. Accordingly, multi-view SSL can be efficiently performed, and the computational complexity can be significantly reduced. We propose an effective algorithm to optimize the objective function. Extensive experimental results on different real-world datasets demonstrate the effectiveness and efficiency of the proposed algorithm.

Learning effective generative models for natural image synthesis is a promising way to reduce the dependence of deep models on massive training data. This work focuses on Fine-Grained Image Synthesis (FGIS) in the semi-supervised setting where a small number of training instances are labeled. Different from generic image synthesis tasks, the available fine-grained data may be inadequate, and the differences among the object categories are typically subtle. To address these issues, we propose a Semantic Regularized class-conditional Generative Adversarial Network, which is referred to as SReGAN. We incorporate an additional discriminator and classifier into the generator-discriminator minimax game. Competing with two discriminators enforces the generator to model both marginal and class-conditional data distributions, which alleviates the problem of limited training data and labels. However, the discriminators may overlook the class separability. To induce the generator to discover the distinctions between classes, we construct semantically congruent and incongruent pairs in the generation process, and further regularize the generator by encouraging high similarities of congruent pairs, while penalizing that of incongruent ones in the classifier's feature space. We have conducted extensive experiments to verify the capability of SReGAN in generating high-fidelity images on a variety of FGIS benchmarks.

Semi-supervised learning is a machine learning approach that integrates supervised and unsupervised learning mechanisms. In this learning, most of labels in the training set are unknown, while there is a small part of data that has known labels. The semi-supervised learning is attractive due to its potential to use labeled and unlabeled data to perform better than supervised learning. This paper consists of a study in the field of semi-supervised learning and implements changes on two well-known semi-supervised learning algorithms: self-training and co-training. In the literature, it is common to develop researches that change the structure of these algorithms, however, none of them proposes automating the labeling process of unlabeled instances, which is the main purpose of this work. In order to achieve this goal, three methods are proposed: FlexCon-G, FlexCon and FlexCon-C. The main difference among these methods is the way in which the confidence rate is calculated and the strategy used to select a label in each iteration. In order to evaluate the proposed methods’ performance, an empirical analysis is conducted, in which the performance of these methods has been evaluated on 30 datasets with different characteristics. The obtained results indicate that all three proposed methods perform better than the original self-training and co-training methods, in most analysed cases.

In the engineering practice, lacking of data especially labeled data typically hinders the wide application of deep learning in mechanical fault diagnosis. However, collecting and labeling data is often expensive and time-consuming. To address this problem, a kind of semi-supervised meta-learning networks (SSMN) with squeeze-and-excitation attention is proposed for few-shot fault diagnosis in this paper. SSMN consists of a parameterized encoder, a non-parameterized prototype refinement process and a distance function. Based on attention mechanism, the encoder is able to extract distinct features to generate prototypes and enhance the identification accuracy. With semi-supervised few-shot learning, SSMN utilizes unlabeled data to refine original prototypes for better fault recognition. A combinatorial learning optimizer is designed to optimize SSMN efficiently. The effectiveness of the proposed method is demonstrated through three bearing vibration datasets and the results indicate the outstanding adaptability in different situations. Comparison with other approaches is also made under the same setup and the experimental results prove the superiority of the proposed method for few-shot fault diagnosis.

We introduce Interpolation Consistency Training (ICT), a simple and computation efficient algorithm for training Deep Neural Networks in the semi-supervised learning paradigm. ICT encourages the prediction at an interpolation of unlabeled points to be consistent with the interpolation of the predictions at those points. In classification problems, ICT moves the decision boundary to low-density regions of the data distribution. Our experiments show that ICT achieves state-of-the-art performance when applied to standard neural network architectures on the CIFAR-10 and SVHN benchmark datasets. Our theoretical analysis shows that ICT corresponds to a certain type of data-adaptive regularization with unlabeled points which reduces overfitting to labeled points under high confidence values., This is the latest version, which is published in the Journal, "Neural Networks", in 2022. All the previous results are unchanged. Keyword: Deep Learning, Semi-supervised Learning, Mixup

Data classification is usually challenged by the difficulty and/or high cost in collecting sufficient labeled data, and unavoidability of data missing. Besides, most of the existing algorithms belong to centralized processing, in which all of the training data must be stored and processed at a fusion center. But in many real applications, data are distributed over multiple nodes, and cannot be centralized to one node for processing due to various reasons. Considering this, in this article, we focus on the problem of distributed classification of missing data with a small proportion of labeled data samples, and develop a distributed semi-supervised missing-data classification (dS²MDC) algorithm. The proposed algorithm is a distributed joint subspace/classifier learning, that is, a latent subspace representation for missing feature imputation is learned jointly with the training of nonlinear classifiers modeled by the χ² kernel using a semi-supervised learning strategy. Theoretical performance analysis and simulations on several datasets clearly validate the effectiveness of the proposed dS²MDC algorithm from different perspectives.

The graph representation (GR) in a data space reveals the intrinsic information as well as the natural relationships of data, which is regarded as a powerful means of representation for solving the semi-supervised learning problem. To effectively learn on a pre-defined graph with both labeled data and unlabeled data, the graph convolutional network (GCN) was proposed and has attracted a lot of attention due to its high-performance graph-based feature extraction along with its low computational complexity. Nevertheless, the performance of GCNs is highly sensitive to the quality of the graph, meaning with high probability the GCNs will achieve poor performances on a badly-defined graphs. In numerous real-world semi-supervised learning problems, the graph connecting each entity in the data space implicitly exists so that there is no naturally pre-defined graph in these problems. To overcome the issues, in this paper, we apply unified graph representation (GR) techniques and graph convolutional (GC) networks in a framework that can be implemented in semi-supervised learning problems. To achieve this framework, we propose sparsity-induced graph convolutional network (SIGCN) for semi-supervised learning. SIGCN introduces the sparsity to formulate significant relationships between instances by constructing a newly-proposed $L_0$ -based graph (termed as the sparsity-induced graph), before applying graph convolution to capture the high-quality features based on this graph for label propagation. We prove and demonstrate the feasibility of the unified framework as well as effectiveness in capturing features. Extensive experiments and comparisons were performed to show the proposed SIGCN obtains a state-of-the-art performance in the semi-supervised learning problem.

When predicting data streams, changes in data distribution may decrease model accuracy over time, thereby making the model obsolete. This phenomenon is known as concept drift. Detecting concept drifts and then adapting to them are critical operations to maintain model performance. However, model adaptation can only be made if labeled data is available. Labeling data is both costly and time-consuming because it has to be done by humans. Only part of the data can be labeled in the data stream because the data size is massive and appears at high speed. To solve these problems simultaneously, we apply a technique to update the model by employing both labeled and unlabeled instances to do so. The experiment results show that our proposed method can adapt to the concept drift with pseudo-labels and maintain its accuracy even though label availability is drastically reduced from 95% to 5%. The proposed method also has the highest overall accuracy and outperforms other methods in 5 of 10 datasets.

Label Propagation Algorithm (LPA) and Graph Convolutional Neural Networks (GCN) are both message passing algorithms on graphs. Both solve the task of node classification, but LPA propagates node label information across the edges of the graph, while GCN propagates and transforms node feature information. However, while conceptually similar, theoretical relationship between LPA and GCN has not yet been systematically investigated. Moreover, it is unclear how LPA and GCN can be combined under a unified framework to improve the performance. Here we study the relationship between LPA and GCN in terms of feature/label influence , in which we characterize how much the initial feature/label of one node influences the final feature/label of another node in GCN/LPA. Based on our theoretical analysis, we propose an end-to-end model that combines GCN and LPA. In our unified model, edge weights are learnable, and the LPA serves as regularization to assist the GCN in learning proper edge weights that lead to improved performance. Our model can also be seen as learning the weights of edges based on node labels, which is more direct and efficient than existing feature-based attention models or topology-based diffusion models. In a number of experiments for semi-supervised node classification and knowledge-graph-aware recommendation, our model shows superiority over state-of-the-art baselines.

PurposeTwitter fake accounts refer to bot accounts created by third-party organizations to influence public opinion, commercial propaganda or impersonate others. The effective identification of bot accounts is conducive to accurately judge the disseminated information for the public. However, in actual fake account identification, it is expensive and inefficient to manually label Twitter accounts, and the labeled data are usually unbalanced in classes. To this end, the authors propose a novel framework to solve these problems.Design/methodology/approachIn the proposed framework, the authors introduce the concept of semi-supervised self-training learning and apply it to the real Twitter account data set from Kaggle. Specifically, the authors first train the classifier in the initial small amount of labeled account data, then use the trained classifier to automatically label large-scale unlabeled account data. Next, iteratively select high confidence instances from unlabeled data to expand the labeled data. Finally, an expanded Twitter account training set is obtained. It is worth mentioning that the resampling technique is integrated into the self-training process, and the data class is balanced at the initial stage of the self-training iteration.FindingsThe proposed framework effectively improves labeling efficiency and reduces the influence of class imbalance. It shows excellent identification results on 6 different base classifiers, especially for the initial small-scale labeled Twitter accounts.Originality/valueThis paper provides novel insights in identifying Twitter fake accounts. First, the authors take the lead in introducing a self-training method to automatically label Twitter accounts from the semi-supervised background. Second, the resampling technique is integrated into the self-training process to effectively reduce the influence of class imbalance on the identification effect.

У роботі розглядаються алгоритми машинного навчання. Увага зосереджена на напівконтрольному навчанні, яке здається балансом між точністю навчання з учителем та витратами методів навчання без учителя. Розглядаються приклади ретельного опрацювання мічених наборів даних, для яких навчання під наглядом може бути дуже ефективним. Порівнюються підходи semi-supervised та supervised та проаналізована ефективність кожного. В роботі розглядаються підходи S3VM та TSVM. Метою роботи було дослідити чи можуть напівконтрольовані підходи конкурувати з контрольованими або навіть їх перевершити. Описується застосування даних підходів до запропонованого датасету для визначення більш точної класифікації даних, а саме на опорній межі.

The popularity of deep learning architecture is increasing day by day. But the majority of deep learning algorithms have their own limitations, such as slow convergence, high training time, sensitivity to noisy data, the problem of local minimum, etc. A deep learning architecture called Multilayer ELM can overcome these limitations to a large extent where there is no backpropagation and hence saves a lot of training time, eliminates the need to fine-tune the parameters, ensure global optimum, able to handle a large volume of data, etc. As above, the most important feature of Multilayer ELM is the characteristics of its feature space, where the input features can linearly be separable without using any kernel techniques, and it still has not taken any attention from the research community. The paper studies the feature space of Multilayer ELM by considering its architecture and feature mapping technique. To justify the effectiveness of its feature mapping technique, semi-supervised and supervised learning algorithms are tested extensively on the feature space of Multilayer ELM and on TF-IDF vector space. Results from the experiment have shown that Multilayer ELM feature space is more effective than the TF-IDF vector space, and the performance of Multilayer ELM is better compared to the baseline machine and deep learning architectures.

In the present work, the study on class imbalance problems in a distributed setting exploiting sparsity structure in the data has been carried out. We formulate the class-imbalance learning problem as a cost-sensitive learning problem with $L_1$ L 1 regularization. The cost-sensitive loss function is a cost-weighted smooth hinge loss. The resultant optimization problem is minimized within the Distributed Alternating Direction Method of Multiplier (DADMM) framework. We partition the data matrix across samples. This operation splits the original problem into a distributed $L_2$ L 2 regularized smooth loss minimization and a $L_1$ L 1 regularized squared loss minimization. $L_2$ L 2 regularized subproblem is solved via Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) and random coordinate descent method in parallel at multiple processing nodes using MPI whereas $L_1$ L 1 regularized problem is just a simple soft-thresholding operation. We show, empirically, that the distributed solution approximates the centralized solution on many benchmark data sets. The centralized solution is obtained via Cost-Sensitive Stochastic Coordinate Descent (CSSCD). Empirical results on small and large-scale benchmark datasets show some promising avenues to further investigate the real-world applications of the proposed algorithms such as anomaly detection, class-imbalance learning, etc. To the best of our knowledge, ours is the first work to study class-imbalance in a distributed environment on large-scale sparse data.

Graph semi-supervised learning (GSSL) plays an important role in data classification by leveraging the similarity across the graph topology and convex optimization with Laplacian-based regularization. However, the current algorithm to solve the problem is centralized approach calling for heavy computational cost, particularly when the data is of large volume. In this paper, an innovate distributed algorithm is proposed to solve the problem, which is based on the decomposition of the similar graph. Contrary to the centralized approach, the distributed algorithm only requires the neighboring information for solving the optimization. It is proved that difference between the solutions of the distributed algorithm and centralized counterpart is upper bounded. We apply the proposed algorithm to both the synthetic and real-world datasets. The numerical results verify the effectiveness of the proposed distributed algorithm.

The semi-supervised learning(SSL) was proposed to deal with the situation that only a few samples were labeled, so how to make the most use of the existing samples is crucial. In general, we apply data augmentation methods, like Generative Adversarial Networks(GAN), to increase the number of data when facing unstructured data. However, things get totally different with structured data, the continuity of neural networks limits their use in category variables. In this paper, we propose a semi-supervised Embedding GAN(EmGAN) to solve that problem. We add an embedding layer before the discriminator to better characterize category features and design a new loss function to further train our model. Moreover, the structures of the generator and discriminator are modified to match structured data. With the experiments of nonparametric statistical test, EmGAN shows its advantages in processing structured samples.

The new coronavirus has caused more than one million deaths and continues to spread rapidly. This virus targets the lungs, causing respiratory distress which can be mild or severe. The X-ray or computed tomography ( CT ) images of lungs can reveal whether the patient is infected with COVID-19 or not. Many researchers are trying to improve COVID-19 detection using artificial intelligence. Our motivation is to develop an automatic method that can cope with scenarios in which preparing labeled data is time consuming or expensive. In this article, we propose a Semi-supervised Classification using Limited Labeled Data ( SCLLD ) relying on Sobel edge detection and Generative Adversarial Networks ( GANs ) to automate the COVID-19 diagnosis. The GAN discriminator output is a probabilistic value which is used for classification in this work. The proposed system is trained using 10,000 CT scans collected from Omid Hospital, whereas a public dataset is also used for validating our system. The proposed method is compared with other state-of-the-art supervised methods such as Gaussian processes. To the best of our knowledge, this is the first time a semi-supervised method for COVID-19 detection is presented. Our system is capable of learning from a mixture of limited labeled and unlabeled data where supervised learners fail due to a lack of sufficient amount of labeled data. Thus, our semi-supervised training method significantly outperforms the supervised training of Convolutional Neural Network ( CNN ) when labeled training data is scarce. The 95% confidence intervals for our method in terms of accuracy, sensitivity, and specificity are 99.56 ± 0.20%, 99.88 ± 0.24%, and 99.40 ± 0.18%, respectively, whereas intervals for the CNN (trained supervised) are 68.34 ± 4.11%, 91.2 ± 6.15%, and 46.40 ± 5.21%.

Defect inspection is an essential part of ensuring the quality of industrial products. Deep learning has achieved great success in defect inspection when a large number of labeled samples are available. However, it is infeasible to collect and label numerous samples in many manufacturing processes. Meanwhile, deep learning methods cannot conform to the high defect recognition accuracy of strict production requirements when the labeled samples are scarce but varied. This paper proposed a novel convolutional neural network architecture and a semi-supervised learning strategy using soft pseudo labels and a mutual correction classifier to improve the defect inspection accuracy when labeled samples are scarce. The effectiveness of the proposed method is verified on a famous industrial defect inspection benchmark dataset and a practical dataset containing images collected from actual injection molding production lines. The results indicate that the proposed method achieves an accuracy of 99.03% on the benchmark defect dataset, which is approximately 13.2% higher than other methods when the training dataset contains only 45 labeled images and 135 unlabeled samples per category. The best accuracy on the benchmark dataset obtained by the proposed method reaches 99.72%. Besides, an average accuracy of 99.25% is achieved with only 20 labeled samples and 180 unlabeled samples per category in the practical defect inspection task. Visualization methods prove that the performance improvement comes from the proposed multiscale architecture and the semi-supervised learning strategy. The proposed method can be used in practical defect inspection applications of industrial manufacturing, such as steel rolling, welding, and injection molding.

In real-world tasks, obtaining a large set of noise-free data can be prohibitively expensive. Therefore, recent research tries to enable machine learning to work with weakly supervised datasets, such as inaccurate or incomplete data. However, the previous literature treats each type of weak supervision individually, although, in most cases, different types of weak supervision tend to occur simultaneously. Therefore, in this article, we present Smart MEnDR, a Classification Model that applies Ensemble Learning and Data-driven Rectification to deal with inaccurate and incomplete supervised datasets. The model first applies a preliminary phase of ensemble learning in which the noisy data points are detected while exploiting the unlabelled data. The phase employs a semi-supervised technique with maximum likelihood estimation to decide on the disagreement rate. Second, the proposed approach applies an iterative meta-learning step to tackle the problem of knowing which points should be made correct to improve the performance of the final classifier. To evaluate the proposed framework, we report the classification performance, noise detection, and the labelling accuracy of the proposed method against state-of-the-art techniques. The experimental results demonstrate the effectiveness of the proposed framework in detecting noise, providing correct labels, and attaining high classification performance.

The label flipping attack is a special poisoning attack in the adversarial environment. The research designed a novel label noise processing framework, the core of which is the semi-supervised learning label correction algorithm based on AdaBoost (AdaSSL). It can effectively improve the label quality of training data and improve the classification performance of the model. Based on five real UCI datasets, this study chose six classic machine learning algorithms (NB, LR, SVM, DT, KNN and MLP) as the base classifiers to classify them. With a noise level of 0 $$ \sim $$ 20%, we evaluated the classification effect of these classifiers on UCI datasets based on the entropy label flipping attack and the AdaSSL defense algorithm. The experimental results show that the AdaSSL algorithm can effectively improve the robustness of the classifier against label flipping attack. Compared with the most advanced semi-supervised defense algorithm in the literature, the algorithm does not need to use additional datasets. At a noise ratio of 10%, the AdaSSL algorithm is significantly better than state-of-the-art label noise defense technology.

At present, computed tomography (CT) is widely used to assist disease diagnosis. Especially, computer aided diagnosis (CAD) based on artificial intelligence (AI) recently exhibits its importance in intelligent healthcare. However, it is a great challenge to establish an adequate labeled dataset for CT analysis assistance, due to the privacy and security issues. Therefore, this paper proposes a convolutional autoencoder deep learning framework to support unsupervised image features learning for lung nodule through unlabeled data, which only needs a small amount of labeled data for efficient feature learning. Through comprehensive experiments, it shows that the proposed scheme is superior to other approaches, which effectively solves the intrinsic labor-intensive problem during artificial image labeling. Moreover, it verifies that the proposed convolutional autoencoder approach can be extended for similarity measurement of lung nodules images. Especially, the features extracted through unsupervised learning are also applicable in other related scenarios.

The performance of supervised deep learning significantly relies on the volume of training samples. However, the vast majority of medical images lacks manual expert annotations. Compared to natural image annotation, the cost of medical image annotation is more expensive as it requires professional medical knowledge guidance. To tackle the predicament, semi-supervised learning and self-supervised learning are very effective technologies. In this paper, we present a twin self-supervision based semi-supervised learning (TS-SSL) approach that embeds two types of self-supervised strategies (namely generative self-supervised learning and discriminative self-supervised learning) into semi-supervised framework to simultaneously learn from few-shot labeled images and vast unlabeled images. TS-SSL is an end-to-end classification model, in which semi-supervision and self-supervision can be jointly trained. The proposed TS-SSL is applied to perform retinal anomaly classification based on spectral-domain optical coherence tomography (SD-OCT) images. The experiments demonstrate that TS-SSL yields the good classification performance on one public SD-OCT dataset and two private SD-OCT datasets with only 10% labels. We also claim that TS-SSL can be transferred to other medical imaging modalities. The code is available at https://github.com/ZhangYH0502/TS-SSL .