Semi-supervised Deep Learning Research Articles

Correcting data imbalance for semi-supervised COVID-19 detection using X-ray chest images

A key factor in the fight against viral diseases such as the coronavirus (COVID-19) is the identification of virus carriers as early and quickly as possible, in a cheap and efficient manner. The application of deep learning for image classification of chest X-ray images of COVID-19 patients could become a useful pre-diagnostic detection methodology. However, deep learning architectures require large labelled datasets. This is often a limitation when the subject of research is relatively new as in the case of the virus outbreak, where dealing with small labelled datasets is a challenge. Moreover, in such context, the datasets are also highly imbalanced, with few observations from positive cases of the new disease. In this work we evaluate the performance of the semi-supervised deep learning architecture known as MixMatch with a very limited number of labelled observations and highly imbalanced labelled datasets. We demonstrate the critical impact of data imbalance to the model’s accuracy. Therefore, we propose a simple approach for correcting data imbalance, by re-weighting each observation in the loss function, giving a higher weight to the observations corresponding to the under-represented class. For unlabelled observations, we use the pseudo and augmented labels calculated by MixMatch to choose the appropriate weight. The proposed method improved classification accuracy by up to 18%, with respect to the non balanced MixMatch algorithm. We tested our proposed approach with several available datasets using 10, 15 and 20 labelled observations, for binary classification (COVID-19 positive and normal cases). For multi-class classification (COVID-19 positive, pneumonia and normal cases), we tested 30, 50, 70 and 90 labelled observations. Additionally, a new dataset is included among the tested datasets, composed of chest X-ray images of Costa Rican adult patients.

Semisupervised dynamic soft sensor based on complementary ensemble empirical mode decomposition and deep learning

Noise, redundancy, and dynamic characteristics in industrial process data have been regarded as the key factors that affect the measurement accuracy of data-driven soft sensors. In this paper, a semi-supervised dynamic soft sensor is proposed to capture the dynamic characteristics of data while removing noise and redundancy within the data, thus ensuring improved accuracy. Complementary ensemble empirical mode decomposition and isometric feature mapping are combined to reduce noise and redundancy. A semi-supervised deep learning model is designed to capture the dynamic characteristics. Compared with traditional soft sensors, the effectiveness and superiority of this method are verified via an experiment using an air preheater of a power boiler. The proposed method achieves the lowest MAE of 0.1745 and the highest correlation coefficient of 0.9969. Compared to methods without data preprocessing, the MAE of the preprocessed soft sensor decreases by 22.28% on average, while the correlation coefficient increases by 0.24% on average.

Increasing the accuracy of single sequence prediction methods using a deep semi-supervised learning framework.

MotivationOver the past 50 years, our ability to model protein sequences with evolutionary information has progressed in leaps and bounds. However, even with the latest deep learning methods, the modelling of a critically important class of proteins, single orphan sequences, remains unsolved.ResultsBy taking a bioinformatics approach to semi-supervised machine learning, we develop Profile Augmentation of Single Sequences (PASS), a simple but powerful framework for building accurate single-sequence methods. To demonstrate the effectiveness of PASS we apply it to the mature field of secondary structure prediction. In doing so we develop S4PRED, the successor to the open-source PSIPRED-Single method, which achieves an unprecedented Q3 score of 75.3% on the standard CB513 test. PASS provides a blueprint for the development of a new generation of predictive methods, advancing our ability to model individual protein sequences.Availability and implementationThe S4PRED model is available as open source software on the PSIPRED GitHub repository (https://github.com/psipred/s4pred), along with documentation. It will also be provided as a part of the PSIPRED web service (http://bioinf.cs.ucl.ac.uk/psipred/).Supplementary information Supplementary data are available at Bioinformatics online.

Building Extraction from Very-High-Resolution Remote Sensing Images Using Semi-Supervised Semantic Edge Detection

The automated detection of buildings in remote sensing images enables understanding the distribution information of buildings, which is indispensable for many geographic and social applications, such as urban planning, change monitoring and population estimation. The performance of deep learning in images often depends on a large number of manually labeled samples, the production of which is time-consuming and expensive. Thus, this study focuses on reducing the number of labeled samples used and proposing a semi-supervised deep learning approach based on an edge detection network (SDLED), which is the first to introduce semi-supervised learning to the edge detection neural network for extracting building roof boundaries from high-resolution remote sensing images. This approach uses a small number of labeled samples and abundant unlabeled images for joint training. An expert-level semantic edge segmentation model is trained based on labeled samples, which guides unlabeled images to generate pseudo-labels automatically. The inaccurate label sets and manually labeled samples are used to update the semantic edge model together. Particularly, we modified the semantic segmentation network D-LinkNet to obtain high-quality pseudo-labels. Specifically, the main network architecture of D-LinkNet is retained while the multi-scale fusion is added in its second half to improve its performance on edge detection. The SDLED was tested on high-spatial-resolution remote sensing images taken from Google Earth. Results show that the SDLED performs better than the fully supervised method. Moreover, when the trained models were used to predict buildings in the neighboring counties, our approach was superior to the supervised way, with line IoU improvement of at least 6.47% and F1 score improvement of at least 7.49%.

Detection and recognition of ultrasound breast nodules based on semi-supervised deep learning: a powerful alternative strategy.

The successful recognition of benign and malignant breast nodules using ultrasound images is based mainly on supervised learning that requires a large number of labeled images. However, because high-quality labeling is expensive and time-consuming, we hypothesized that semi-supervised learning could provide a low-cost and powerful alternative approach. This study aimed to develop an accurate semi-supervised recognition method and compared its performance with supervised methods and sonographers. The faster region-based convolutional neural network was used for nodule detection from ultrasound images. A semi-supervised classifier based on the mean teacher model was proposed to recognize benign and malignant nodule images. The general performance of the proposed method on two datasets (8,966 nodules) was reported. The detection accuracy was 0.88±0.03 and 0.86±0.02, respectively, on two testing sets (1,350 and 2,220 nodules). When 800 labeled training nodules were available, the proposed semi-supervised model plus 4,396 unlabeled nodules performed better than the supervised learning model (area under the curve (AUC): 0.934±0.026 vs. 0.83±0.050; 0.916±0.022 vs. 0.815±0.049). The performance of the semi-supervised model trained on 800 labeled and 4,396 unlabeled nodules was close to that of the supervised learning model trained on a massive number of labeled nodules (n=5,196) (AUC: 0.934±0.026 vs. 0.952±0.027; 0.916±0.022 vs. 0.918±0.017). Moreover, the semi-supervised model was better than the average accuracy of five human sonographers (AUC: 0.922 vs. 0.889). The semi-supervised model can achieve excellent performance for nodule recognition and be useful for medical sciences. The method reduced the number of labeled images required for training, thus significantly alleviating the difficulty in data preparation of medical artificial intelligence.

A semi-supervised deep learning approach for predicting the functional effects of genomic non-coding variations

BackgroundUnderstanding the functional effects of non-coding variants is important as they are often associated with gene-expression alteration and disease development. Over the past few years, many computational tools have been developed to predict their functional impact. However, the intrinsic difficulty in dealing with the scarcity of data leads to the necessity to further improve the algorithms. In this work, we propose a novel method, employing a semi-supervised deep-learning model with pseudo labels, which takes advantage of learning from both experimentally annotated and unannotated data.ResultsWe prepared known functional non-coding variants with histone marks, DNA accessibility, and sequence context in GM12878, HepG2, and K562 cell lines. Applying our method to the dataset demonstrated its outstanding performance, compared with that of existing tools. Our results also indicated that the semi-supervised model with pseudo labels achieves higher predictive performance than the supervised model without pseudo labels. Interestingly, a model trained with the data in a certain cell line is unlikely to succeed in other cell lines, which implies the cell-type-specific nature of the non-coding variants. Remarkably, we found that DNA accessibility significantly contributes to the functional consequence of variants, which suggests the importance of open chromatin conformation prior to establishing the interaction of non-coding variants with gene regulation.ConclusionsThe semi-supervised deep learning model coupled with pseudo labeling has advantages in studying with limited datasets, which is not unusual in biology. Our study provides an effective approach in finding non-coding mutations potentially associated with various biological phenomena, including human diseases.

A survey on deep learning models for wireless capsule endoscopy image analysis

Abdomen Bleeding, Ulcer, Tumour, Crohn's disease, Celiac disease and other diseases in the gastrointestinal tract (GI) are difficult to diagnose, due to the inescapable inherent difficulty in accessing a volute setting in the human body. Wireless Capsule Endoscopy (WCE) offers a patient-cordial, non-invasive and painless investigation in the GI tract. Automatic detection of anomalies in WCE images using Deep Learning Models improves the detection accuracy but it requires a huge number of labeled data for model training. But these deep models suffer from explain-ability and fail to include expert knowledge in the model decision-making process. By keeping these aspects in mind, this survey aims to identify the opportunities for using Semi-Supervised deep learning models over supervised deep learning methods in Wireless Capsule Endoscopy (WCE) anomaly detection and classification. This paper presents a comprehensive survey on various deep learning solutions for anomaly detection and localization techniques utilized in WCE images in the aspect of performance, complexity, and the quality of the dataset. The survey outlined the proposed Attention and Domain Assisted Generative Adversarial Network (ADA-GAN) based Semi-Supervised Model for WCE anomaly image classification along with initial results. The result derives the shortcomings of the current literature methods and paves the potential research opportunities in the Semi-Supervised models in Wireless capsule endoscopy image analysis.

DeHiB: Deep Hidden Backdoor Attack on Semi-supervised Learning via Adversarial Perturbation

The threat of data-poisoning backdoor attacks on learning algorithms typically comes from the labeled data. However, in deep semi-supervised learning (SSL), unknown threats mainly stem from the unlabeled data. In this paper, we propose a novel deep hidden backdoor (DeHiB) attack scheme for SSL-based systems. In contrast to the conventional attacking methods, the DeHiB can inject malicious unlabeled training data to the semi-supervised learner so as to enable the SSL model to output premeditated results. In particular, a robust adversarial perturbation generator regularized by a unified objective function is proposed to generate poisoned data. To alleviate the negative impact of the trigger patterns on model accuracy and improve the attack success rate, a novel contrastive data poisoning strategy is designed. Using the proposed data poisoning scheme, one can implant the backdoor into the SSL model using the raw data without hand-crafted labels. Extensive experiments based on CIFAR10 and CIFAR100 datasets demonstrated the effectiveness and crypticity of the proposed scheme.

Fuzzy Integration Algorithm of Big Data in Peer-to-Peer Communication Network Based on Deep Learning

There are many kinds of big data in peer to peer communication network, traditional big data integration algorithms ignore the classification of big data state. The location correlation between big data cannot be accurately obtained, which leads to the time-consuming process of big data integration and the low acquisition accuracy of big data location correlation. A fuzzy integration algorithm for big data in peer-to-peer communication network based on deep learning is proposed. According to the big data integration conditions of peer-to-peer communication network, the benchmark model of big data integration is constructed by combining semi-supervised deep learning. The marked sample updating model of the prior single classification fuzzy integrated data model is used to process the corresponding sample data. Combining the steady-state and dynamic residuals of the datum model, multiple information grains are obtained. Based on this, kalman filter is adopted to fuzzy fusion of big data to obtain state parameters, and multi-scale analysis is carried out to filter out big data noise. Using the obtained information grains to constrain the characteristic correlation degree of big data, the position correlation of big data is completed. Implement fuzzy integration algorithm design of big data in peer-to-peer communication network based on deep learning. In order to verify the effectiveness of the proposed algorithm, a simulation experiment is designed. Experimental results show that compared with many traditional methods, the proposed algorithm takes less time and the computational complexity of the proposed algorithm is significantly reduced. And the big data integration is more accurate and has better application effect.

Driving Maneuvers Detection using Semi-Supervised Long Short-Term Memory and Smartphone Sensors

Driving maneuvers detection is an important component of proactive traffic safety management and connected vehicle systems. Most of the existing studies used supervised learning concepts to train their models with labeled data. These methods achieved promising results but were limited by the heavy dependence on the labeled data. With the development of mobile sensing technologies, massive traffic-related data can be efficiently collected by mobile devices (e.g., smartphones, tablets, etc.). Considering the high costs of labeling data, this paper proposed a semi-supervised deep learning method to learn from the unlabeled data. Data from a smartphone’s accelerometer and gyroscope were collected by different drivers with a variety of smartphones, vehicles, and locations. Three long short-term memory (LSTM) models were trained with the proposed semi-supervised learning algorithm. Experimental results indicated that the proposed semi-supervised LSTM could learn from the unlabeled data and achieve outstanding results with only a small portion of the labeled data. Using much fewer labeled data, semi-supervised LSTM could achieve similar results compared with the supervised method. Moreover, the proposed method outperformed other machine learning methods (e.g., convolutional neural network, XGBoost, random forest) on precision, recall, F1-score, and area under curve. More and more traffic data will be available in the future, the proposed method is expected to make use of the undiscovered potential from the massive unlabeled data.

Cost-sensitive semi-supervised deep learning to assess driving risk by application of naturalistic vehicle trajectories

Most traffic accidents are caused by driver-related factors such as poor perception, aggressive decision-making, or improper maneuvering. Therefore, it is critical to evaluate and predict driving risks to provide drivers with timely feedback. However, risk assessment involves challenges related to a lack of labeled driving data and the presence of data imbalance in the description of different driving risk levels. To address these challenges, a cost-sensitive semi-supervised deep learning method is developed to obtain driving risk scores based on naturalistic vehicle trajectories. A convolutional neural network and a long short-term memory encoder/decoder network are embedded into a semi-supervised framework that uses only a small labeled dataset to label the remaining unlabeled data and produce a trained network model. As fixed weights cannot adapt to changes in the degree of class imbalance that occur over progressive semi-supervised learning iterations, an adaptive over-balanced cross-entropy loss function is developed to adaptively maintain an over-balanced state for the high-risk class to achieve cost-sensitive learning. The experimental results indicate that the accuracy of the proposed method in determining the current and future 2 s risk scores is 96.63% and 92.06%, respectively, thereby constituting the best comprehensive performance among existing machine learning methods. Moreover, the method is verified using a spatio-temporal diagram of driving risk-trajectory and a current–future risk score diagram. The findings demonstrate that the proposed method can be used to assess driving risks in a reliable and robust manner.

Evaluation of structural integrity of railway bridge using acceleration data and semi-supervised learning approach

In order to investigate the structural integrity of a railway bridge, supervised and semi-supervised deep learning techniques incorporating wavelet transform (WT) were used where the measured acceleration data were transformed into images. Specifically, the structural integrity was evaluated both by acceleration datasets of a newly built bridge and by damaged datasets calculated from numerical analyses, and assumed damage in the structure was detected by supervised and semi-supervised learning methods. For the supervised learning, the well-known AlexNet and VGG16 convolutional neural network (CNN) models were employed, and during the one class (OC) classification of the semi-supervised learning approach, wherein only the label of the normal data was known a priori, a transfer learning method was utilized. It was found that the minimum value of damage with which novelty was detected was found to be at least 15% reduction of the stiffness in our case. It was also found that the cosine rather than Euclidean distance metric was more accurate during damage prediction. Although both methods showed very reliable prediction results, the semi-supervised learning method was shown to be more practical under the given field conditions.

Whole‐brain functional MRI registration based on a semi‐supervised deep learning model

Traditional registration of functional magnetic resonance images (fMRI) is typically achieved through registering their coregistered structural MRI. However, it cannot achieve accurate performance in that functional units which are not necessarily located relative to anatomical structures. In addition, registration methods based on functional information focus on gray matter (GM) information but ignore the importance of white matter (WM). To overcome the limitations of exiting techniques, in this paper, we aim to register resting-state fMRI (rs-fMRI) based directly on rs-fMRI data and make full use of GM and WM information to improve the registration performance. We provide a robust representation of WM functional connectivity features using tissue-specific patch-based functional correlation tensors (ts-PFCTs) as auxiliary information to assist registration. Furthermore, we propose a semi-supervised deep learning model that uses GM and WM information (GM ts-PFCTs and WM ts-PFCTs) during training as a fine tweak to improve registration accuracy when such information is not provided in new test image pairs. We implement our method on the 1000 Functional Connectomes Project dataset. To evaluate our method, a group-level analysis was implemented in resting-state brain functional networks after registration, resulting in t maps. Our method increases the peak t values of the t maps of default mode network, visual network, central executive network, and sensorimotor network to 21.4, 20.0, 18.4, and 19.0, respectively. Through comparison with traditional methods (FMRIB Software Library(FSL), Statistical Parametric Mapping _ Echo Planar Image(SPM_EPI), and SPM_T1), our method achieves an average improvement of 67.39%, 12.96%, and 25.14%. We propose a semi-supervised deep learning network by adding GM and WM information as auxiliary information for resting-state fMRI registration. GM and WM information is extracted and described as GM ts-PFCTs and WM ts-PFCTs. Experimental results show that our method achieves superior registration performance.

Automatic detection of retinopathy with optical coherence tomography images via a semi-supervised deep learning method.

Automatic detection of retinopathy via computer vision techniques is of great importance for clinical applications. However, traditional deep learning based methods in computer vision require a large amount of labeled data, which are expensive and may not be available in clinical applications. To mitigate this issue, in this paper, we propose a semi-supervised deep learning method built upon pre-trained VGG-16 and virtual adversarial training (VAT) for the detection of retinopathy with optical coherence tomography (OCT) images. It only requires very few labeled and a number of unlabeled OCT images for model training. In experiments, we have evaluated the proposed method on two popular datasets. With only 80 labeled OCT images, the proposed method can achieve classification accuracies of 0.942 and 0.936, sensitivities of 0.942 and 0.936, specificities of 0.971 and 0.979, and AUCs (Area under the ROC Curves) of 0.997 and 0.993 on the two datasets, respectively. When comparing with human experts, it achieves expert level with 80 labeled OCT images and outperforms four out of six experts with 200 labeled OCT images. Furthermore, we also adopt the Gradient Class Activation Map (Grad-CAM) method to visualize the key regions that the proposed method focuses on when making predictions. It shows that the proposed method can accurately recognize the key patterns of the input OCT images when predicting retinopathy.

Semi-supervised deep learning approach to break common CAPTCHAs

Manual data annotation is a time consuming activity. A novel strategy for automatic training of the CAPTCHA breaking system with no manual dataset creation is presented in this paper. We demonstrate the feasibility of the attack against a text-based CAPTCHA scheme utilizing similar network infrastructure used for Denial of Service attacks. The main goal of our research is to present a possible vulnerability in CAPTCHA systems when combining the brute-force attack with transfer learning. The classification step utilizes a simple convolutional neural network with 15 layers. Training stage uses automatically prepared dataset created without any human intervention and transfer learning for fine-tuning the deep neural network classifier. The designed system for breaking text-based CAPTCHAs achieved 80% classification accuracy after 6 fine-tuning steps for a 5 digit text-based CAPTCHA system. The results presented in this paper suggest, that even the simple attack with a large number of attacking computers can be an effective alternative to current CAPTCHA breaking systems.

Insecurity Early Warning for Large Scale Hybrid AC/DC Grids Based on Decision Tree and Semi-Supervised Deep Learning

Fast insecurity early warning is the key technique to resist the dynamic insecurity risk, which becomes intractable due to the strong nonlinearity of hybrid AC/DC grids and the high uncertainty of wind generation. Considering dynamic security constraints and wind power uncertainty, this paper presents an insecurity early warning method based on decision tree (DT) and semi-supervised deep learning. First, semi-supervised deep learning is deployed to estimate the dynamic security limit of the critical interface of hybrid AC/DC grids. The system security is assessed by comparing the actual power transfer of the critical interface with the security limit. Then, operating conditions (OCs) are ranked into different insecure levels according to the type of preventive control actions that is needed to ensure the system security. Finally, oblique DT is utilized to identify insecurity classification boundaries in the wind power injection space. Insecure OC sets are constructed based on these classification boundaries. Simulation results of the real-life Jiangsu-Shanghai interconnected grid in east China demonstrate that the proposed method can fast construct the insecure OC sets corresponding to different insecure levels.

SeBioGraph: Semi-supervised Deep Learning for the Graph via Sustainable Knowledge Transfer.

Semi-supervised deep learning for the biomedical graph and advanced manufacturing graph is rapidly becoming an important topic in both academia and industry. Many existing types of research focus on semi-supervised link prediction and node classification, as well as the application of these methods in sustainable development and advanced manufacturing. To date, most manufacturing graph neural networks are mainly evaluated on social and information networks, which improve the quality of network representation y integrating neighbor node descriptions. However, previous methods have not yet been comprehensively studied on biomedical networks. Traditional techniques fail to achieve satisfying results, especially when labeled nodes are deficient in number. In this paper, a new semi-supervised deep learning method for the biomedical graph via sustainable knowledge transfer called SeBioGraph is proposed. In SeBioGraph, both node embedding and graph-specific prototype embedding are utilized as transferable metric space characterized. By incorporating prior knowledge learned from auxiliary graphs, SeBioGraph further promotes the performance of the target graph. Experimental results on the two-class node classification tasks and three-class link prediction tasks demonstrate that the SeBioGraph realizes state-of-the-art results. Finally, the method is thoroughly evaluated.

Semi-supervised deep learning of source location and seabed class from unlabeled merchant ships in the shallow ocean

Understanding sound propagation in a shallow ocean environment is complicated but important for naval technologies. Some researchers are currently applying a variety of supervised deep learning methods, such as convolution neural networks, for either source localization or seabed classification. One of the costs of supervised learning is the requirement for large amounts of labeled data. In the present study, unlabeled data are used to train a Transformer model. Transformer-based models have demonstrated ability for predicting sequential data; examples are Google’s BERT and OpenAI’s GPT-3. In our technique, we train a transformer-based model in a twopart process. First, self-supervised learning is implemented using synthetic ship spectrograms for various shallow ocean environments. The model is trained as an encoder/decoder to perform sequence-to-sequence prediction. Second, the transformer model is fine-tuned to predict the source location and seabed class from a small set of labeled synthetic samples. Data samples measured during the Seabed Characterization Experiment 2017 are utilized as a testing dataset. The advantages of this approach are the ability to train a model with a larger variety of data, including the use of unlabeled data and data of variable input length.

Multi-modal generative adversarial networks for traffic event detection in smart cities

Advances in the Internet of Things have enabled the development of many smart city applications and expert systems that help citizens and authorities better understand the dynamics of the cities, and make better planning and utilisation of city resources. Smart cities are composed of complex systems that usually process and analyse big data from the Cyber, Physical, and Social worlds. Traffic event detection is an important and complex task in smart transportation modelling and management. We address this problem using semi-supervised deep learning with data of different modalities, e.g., physical sensor observations and social media data. Unlike most existing studies focusing on data of single modality, the proposed method makes use of data of multiple modalities that appear to complement and reinforce each other. Meanwhile, as the amount of labelled data in big data applications is usually extremely limited, we extend the multi-modal Generative Adversarial Network model to a semi-supervised architecture to characterise traffic events. We evaluate the model with a large, real-world dataset consisting of traffic sensor observations and social media data collected from the San Francisco Bay Area over a period of four months. The evaluation results clearly demonstrate the advantages of the proposed model in extracting and classifying traffic events.

Machine seismic: an automatic approach for the identification of subsurface structural models

Subsurface structure identification and modeling in three-dimensional dataset is a very complex problem for the geoscientist. Manual dealing of seismic data involves different complexities and requires more time and effort in order to address the errors and mistakes associated with the quality of seismic data. It is sometimes not straightforward to classify the features of interest especially when the data are of poor quality or associated with complex micro-tectonic regime. This brings the need of automatic working engine which can handle such issues and can find the target-oriented features with the more accuracy. Therefore, advance machine learning approach has been considered as a powerful tool which extensively used in different fields of data science to enhance the quality of data and fast processing. In this study, a new technology has been introduced in the seismic processing and development to make processes in automated routine. This method is based on deep neural network (DNN) that directly translates the raw seismic data into final fault models. A test example of deep semi-supervised learning method was performed for seismic faults and reflector identification without prior information and without going into complete dataset for labeling. This example explains the challenges to the issue and gives the solution to the problem related to the structure and tectonic modeling to explain the performance gain. In this approach of semi-supervised learning, only one of any reflectors and one or two of the faults has been introduced in the learning phase to identify such features in the complete 3D cube. This is an example of deep learning neural network for machine seismic in real case study and allows exploration geoscientists to easier mark the hot spots of their interest before mining or drilling.