Unlabeled Dataset Research Articles

Content-Based Image Retrieval Using Deep Learning

The most prevalent and well-used method for obtaining images from huge, unlabelled image datasets is content-based image retrieval. Convolutional Neural Networks are pre-trained deep neural networks which can generate and extract accurate features from image databases. These CNN models have been trained using large databases with thousands of classes that include a huge number of images, making it simple to use their information. Based on characteristics retrieved using the pre-trained CNN models, we created CBIR systems in the work. These pre-trained CNN models VGG16, and MobileNet have been employed in this instance to extract sets of features that are afterward saved independently and used for image retrieval.

MLDRL: Multi-loss disentangled representation learning for predicting esophageal cancer response to neoadjuvant chemoradiotherapy using longitudinal CT images

Accurate prediction of pathological complete response (pCR) after neoadjuvant chemoradiotherapy (nCRT) is essential for clinical precision treatment. However, the existing methods of predicting pCR in esophageal cancer are based on the single stage data, which limits the performance of these methods. Effective fusion of the longitudinal data has the potential to improve the performance of pCR prediction, thanks to the combination of complementary information. In this study, we propose a new multi-loss disentangled representation learning (MLDRL) to realize the effective fusion of complementary information in the longitudinal data. Specifically, we first disentangle the latent variables of features in each stage into inherent and variational components. Then, we define a multi-loss function to ensure the effectiveness and structure of disentanglement, which consists of a cross-cycle reconstruction loss, an inherent-variational loss and a supervised classification loss. Finally, an adaptive gradient normalization algorithm is applied to balance the training of multiple loss terms by dynamically tuning the gradient magnitudes. Due to the cooperation of the multi-loss function and the adaptive gradient normalization algorithm, MLDRL effectively restrains the potential interference and achieves effective information fusion. The proposed method is evaluated on multi-center datasets, and the experimental results show that our method not only outperforms several state-of-art methods in pCR prediction, but also achieves better performance in the prognostic analysis of multi-center unlabeled datasets.

A Deep Learning Powered System to Lie Detection While Online Study

Many education facilities have recently switched to online learning due to the COVID-19 pandemic. The nature of online learning makes it easier for dishonest behaviors, such as cheating or lying during lessons. We propose a new artificial intelligence - powered solution to help educators solve this rising problem for a fairer learning environment. We created a visual representation contrastive learning method with the MobileNetV2 network as the backbone to improve predictability from an unlabeled dataset which can be deployed on low power consumption devices. The experiment shows an accuracy of up to 59%, better than several previous research, proving the usability of this approach.

Suggestive Data Annotation for CNN-Based Building Footprint Mapping Based on Deep Active Learning and Landscape Metrics

Convolutional neural network (CNN)-based very high-resolution (VHR) image segmentation has become a common way of extracting building footprints. Despite publicly available building datasets and pre-trained CNN models, it is still necessary to prepare sufficient labeled image tiles to train CNN models from scratch or update the parameters of pre-trained CNN models to extract buildings accurately in real-world applications, especially the large-scale building extraction, due to differences in landscapes and data sources. Deep active learning is an effective technique for resolving this issue. This study proposes a framework integrating two state-of-the-art (SOTA) models, U-Net and DeeplabV3+, three commonly used active learning strategies, (i.e., margin sampling, entropy, and vote entropy), and landscape characterization to illustrate the performance of active learning in reducing the effort of data annotation, and then understand what kind of image tiles are more advantageous for CNN-based building extraction. The framework enables iteratively selecting the most informative image tiles from the unlabeled dataset for data annotation, training the CNN models, and analyzing the changes in model performance. It also helps us to understand the landscape features of iteratively selected image tiles via active learning by considering building as the focal class and computing the percent, the number of patches, edge density, and landscape shape index of buildings based on labeled tiles in each selection. The proposed method was evaluated on two benchmark building datasets, WHU satellite dataset II and WHU aerial dataset. Models in each iteration were trained from scratch on all labeled tiles. Experimental results based on the two datasets indicate that, for both U-Net and DeeplabV3+, the three active learning strategies can reduce the number of image tiles to be annotated and achieve good model performance with fewer labeled image tiles. Moreover, image tiles with more building patches, larger areas of buildings, longer edges of buildings, and more dispersed building distribution patterns were more effective for model training. The study not only provides a framework to reduce the data annotation efforts in CNN-based building extraction but also summarizes the preliminary suggestions for data annotation, which could facilitate and guide data annotators in real-world applications.

High-fidelity positive-unlabeled deep learning for semi-supervised fault detection of chemical processes

With the rapid development of the modern chemical process industry, process monitoring techniques have been investigated to enhance the loss prevention capability. Fault detection which attempts to detect whether an abnormal condition has happened is an essential step of process monitoring. Supervised learning based fault detection methods usually are not feasible in practical industry situations, because their most important prerequisite that adequate data labels are accessible is extremely hard to meet. In real industrial situations, a common scenario is that a few normal state data are labeled while no fault state data are labeled, and this is a natural fit for the positive-unlabeled (PU) learning. In this work, a three-step high-fidelity PU (THPU) approach based on deep learning is proposed for semi-supervised fault detection of chemical processes. A self-training nearest neighbors (STNN) algorithm is designed in Step 1 to conservatively generate a small reliable positive dataset by extracting data from the unlabeled training dataset. A positive and negative data recognition (PNDR) algorithm is designed in Step 2 to augment the reliable positive dataset and generate a reliable negative dataset based on the remaining unlabeled dataset. A convolutional neural network is trained in Step 3 based on the above reliable datasets as the binary classifier for online fault detection. The benchmark simulated Tennessee Eastman process dataset and a real industrial hydro-cracking process dataset are utilized to illustrate the effectiveness and robustness of the proposed THPU approach. The experimental results demonstrate the superiority of the proposed approach compared to the other competing PU learning approaches and supervised fault detection models.

Self-Supervised Pretraining for RGB-D Salient Object Detection

Existing CNNs-Based RGB-D salient object detection (SOD) networks are all required to be pretrained on the ImageNet to learn the hierarchy features which helps provide a good initialization. However, the collection and annotation of large-scale datasets are time-consuming and expensive. In this paper, we utilize self-supervised representation learning (SSL) to design two pretext tasks: the cross-modal auto-encoder and the depth-contour estimation. Our pretext tasks require only a few and unlabeled RGB-D datasets to perform pretraining, which makes the network capture rich semantic contexts and reduce the gap between two modalities, thereby providing an effective initialization for the downstream task. In addition, for the inherent problem of cross-modal fusion in RGB-D SOD, we propose a consistency-difference aggregation (CDA) module that splits a single feature fusion into multi-path fusion to achieve an adequate perception of consistent and differential information. The CDA module is general and suitable for cross-modal and cross-level feature fusion. Extensive experiments on six benchmark datasets show that our self-supervised pretrained model performs favorably against most state-of-the-art methods pretrained on ImageNet. The source code will be publicly available at https://github.com/Xiaoqi-Zhao-DLUT/SSLSOD.

XDC: Adversarial Adaptive Cross Domain Face Clustering (Student Abstract)

In this work we propose a scheme, called XDC, that uses adversarial learning to train an adaptive cross domain clustering model. XDC trains a classifier on a labeled dataset and assigns labels to an unlabeled dataset. We benefit from adversarial learning such that the target dataset takes part in the training. We also use an existing image classifiers in a plug-and-play fashion (i.e, it can be replaced with any other image classifier). Unlike existing works we update the parameters of the encoder and expose the target dataset to the model during training. We apply our model on two face dataset and one non-face dataset and obtain comparable results with state-of-the-art face clustering models.

RareGAN: Generating Samples for Rare Classes

We study the problem of learning generative adversarial networks (GANs) for a rare class of an unlabeled dataset subject to a labeling budget. This problem is motivated from practical applications in domains including security (e.g., synthesizing packets for DNS amplification attacks), systems and networking (e.g., synthesizing workloads that trigger high resource usage), and machine learning (e.g., generating images from a rare class). Existing approaches are unsuitable, either requiring fully-labeled datasets or sacrificing the fidelity of the rare class for that of the common classes. We propose RareGAN, a novel synthesis of three key ideas: (1) extending conditional GANs to use labelled and unlabelled data for better generalization; (2) an active learning approach that requests the most useful labels; and (3) a weighted loss function to favor learning the rare class. We show that RareGAN achieves a better fidelity-diversity tradeoff on the rare class than prior work across different applications, budgets, rare class fractions, GAN losses, and architectures.

Positive-unlabeled classification under class-prior shift: a prior-invariant approach based on density ratio estimation

Learning from positive and unlabeled (PU) data is an important problem in various applications. Most of the recent approaches for PU classification assume that the class-prior (the ratio of positive samples) in the training unlabeled dataset is identical to that of the test data, which does not hold in many practical cases. In addition, we usually do not know the class-priors of the training and test data, thus we have no clue on how to train a classifier without them. To address these problems, we propose a novel PU classification method based on density ratio estimation. A notable advantage of our proposed method is that it does not require the class-priors in the training phase; class-prior shift is incorporated only in the test phase. We theoretically justify our proposed method and experimentally demonstrate its effectiveness.

Low-budget label query through domain alignment enforcement

Deep learning revolution happened thanks to the availability of a massive amount of labeled data which contributed to the development of models with extraordinary inference capabilities. Despite the public availability of large-scale datasets, to address specific requirements it is often necessary to generate a new set of labeled data whose production is often costly and require specific know-how to be fulfilled. In this work, we propose the new problem of low-budget label query, which aims at maximizing the classification performance by selecting a convenient and small set of samples (i.e., low budget) to be manually labeled from an arbitrary big set of unlabeled data. While a first solution might be the use of pre-trained models with standard selection metrics, i.e. confidence and entropy, we argue that domain shift affects their reliability. We deem that Unsupervised Domain Adaptation (UDA) can be used to reduce domain shift, making selection metrics more reliable and less noisy. Therefore, we first improve an UDA method to better align source and target domains using consistency constraints, reaching comparable performance with the state of-the-art on several UDA tasks. After adaptation, we conduct an extensive experimental study with commonly used confidence metrics and sampling strategies to achieve low-budget label query on a large variety of publicly available datasets and under different setups.

Ontology based sentiment analysis for fake review detection

Majority of customers and manufacturers who tend to purchase and trade via e-commerce websites primarily rely on reviews before making purchasing decisions and product improvements. Deceptive reviewers consider this opportunity to write fake reviews to mislead customers and manufacturers. This calls for the necessity of identifying fake reviews before making them available for decision making. Accordingly, this research focuses on a fake review detection method that incorporates review-related features including linguistic features, Part-of-Speech (POS) features, and sentiment analysis features. A domain feature ontology is used in the feature-level sentiment analysis and all the review-related features are extracted and integrated into the ontology. The fake review detection is enhanced through a rule-based classifier by inferencing the ontology. Due to the lack of a labeled dataset for model training, the Mahalanobis distance method was used to detect outliers from an unlabeled dataset where the outliers were selected as fake reviews for model training. The performance measures of the rule-based classifier were improved by integrating linguistic features, POS features, and sentiment analysis features, in spite of considering them separately.

Incorporating Measurement Error in Astronomical Object Classification

Most general-purpose classification methods, such as support-vector machine (SVM) and random forest (RF), fail to account for an unusual characteristic of astronomical data: known measurement error uncertainties. In astronomical data, this information is often given in the data but discarded because popular machine learning classifiers cannot incorporate it. We propose a simulation-based approach that incorporates heteroscedastic measurement error into an existing classification method to better quantify uncertainty in classification. The proposed method first simulates perturbed realizations of the data from a Bayesian posterior predictive distribution of a Gaussian measurement error model. Then, a chosen classifier is fit to each simulation. The variation across the simulations naturally reflects the uncertainty propagated from the measurement errors in both labeled and unlabeled data sets. We demonstrate the use of this approach via two numerical studies. The first is a thorough simulation study applying the proposed procedure to SVM and RF, which are well-known hard and soft classifiers, respectively. The second study is a realistic classification problem of identifying high-z (2.9 ≤ z ≤ 5.1) quasar candidates from photometric data. The data are from merged catalogs of the Sloan Digital Sky Survey, the Spitzer IRAC Equatorial Survey, and the Spitzer-HETDEX Exploratory Large-Area Survey. The proposed approach reveals that out of 11,847 high-z quasar candidates identified by a random forest without incorporating measurement error, 3146 are potential misclassifications with measurement error. Additionally, out of 1.85 million objects not identified as high-z quasars without measurement error, 936 can be considered new candidates with measurement error.

Active Learning Query Strategies for Linear Regression Based on Efficient Global Optimization

Active learning, a subfield of machine learning, can train a good model by selecting a minimum number of labeled samples. In many machine learning scenarios, needed information (such as the best value in unlabeled datasets) is acquired by prediction. When there is too little data in the training model, the prediction accuracy would obviously affect the accuracy of the results. To establish a high-performance regression model for a small dataset while accelerating the search for the best sample, a new active learning query strategy, EGO-ALR, that combines efficient global optimization (EGO) and active learning for regression (ALR) was proposed. It was found that the performance of EGO-ALR was significantly better than the original ALR query strategies in terms of the root mean square error (RMSE), correlation coefficient (CC), and opportunity cost (Oppo Cost). Specifically, EGO-ALR increased the Oppo Cost by an average of 25.27% when the RMSE and CC values were not more than 1.07% different from the original ALR. This study validated the efficiency and robustness of EGO-ALR approaches using 19 datasets from various domains and three distinct linear regression models (Ridge regression, Lasso, and Elastic network).

Unsupervised contrastive learning for few-shot TOC prediction and application

Total organic carbon (TOC) content is significant for “sweet spots” prediction and resource calculation. Traditional physical methods and machine learning methods, however, only use labeled samples (core TOC samples and well logging data) to predict TOC, resulting in a performance bottleneck. Meanwhile, many oilfields have considerable unlabeled data (well logging data). Therefore, this paper proposed a pretrain-and-finetune few-shot learning framework, called Contrastive Learning-Convolutional Neural Network (CL-CNN), to predict TOC using unlabeled data. CL has been a research hotspot in unsupervised learning in recent years, with the principle of reducing the distance between similar samples and increasing the distance between different samples in the feature space. CNN is distinguished by its strong generalization ability and high prediction accuracy, as well as its local connectivity and weight sharing. The CL-CNN model first uses CL to pretrain the initial parameters of the CNN model on a large number of unlabeled datasets and then finetunes the resultant CNN model using the labeled data. The framework combines the advantages of CL and CNN to further improve the prediction accuracy and generalization ability. The proposed method was applied to predict the TOC of shale gas in the Sichuan Basin, China, using 121 core samples and 1224 well logging data from the JY 10 well (121 logging data for supervised learning and 1103 logging data for unsupervised learning). Besides, JY 10–10 well data were used to validate the CL-CNN model. Experimental results demonstrated that the CL-CNN model outperformed many machine learning methods, including the support vector machine (SVM), Gaussian process regression (GPR), random forest (RF), and CNN models, indicating the feasibility and effectiveness of the proposed method. Because of this, the CL-CNN model can also be used as a general prediction tool for reservoir parameters such as porosity and permeability.

Source-free unsupervised domain adaptation for cross-modality abdominal multi-organ segmentation

Domain adaptation is crucial for transferring the knowledge from the source labeled CT dataset to the target unlabeled MR dataset in abdominal multi-organ segmentation. Meanwhile, it is highly desirable to avoid the high annotation cost related to the target dataset and protect the source dataset privacy. Therefore, we propose an effective source-free unsupervised domain adaptation method for cross-modality abdominal multi-organ segmentation without source dataset access. The proposed framework comprises two stages. In the first stage, the feature map statistics-guided model adaptation combined with entropy minimization is developed to help the top segmentation network reliably segment the target images. The pseudo-labels output from the top segmentation network are used to guide the style compensation network to generate source-like images. The pseudo-labels output from the middle segmentation network is used to supervise the learning progress of the desired model (bottom segmentation network). In the second stage, the circular learning and pixel-adaptive mask refinement are used to further improve the desired model performance. With this approach, we achieved satisfactory abdominal multi-organ segmentation performance, outperforming the existing state-of-the-art domain adaptation methods. The proposed approach can be easily extended to situations in which target annotation data exist. With only one labeled MR volume, the performance can be leveled with that of supervised learning. Furthermore, the proposed approach is proven to be effective for source-free unsupervised domain adaptation in reverse direction.

An Unsupervised Software Fault Prediction Approach Using Threshold Derivation

Software fault prediction models help the software quality assurance team to manage the resources optimally during software maintenance. Most of the recently proposed fault prediction approaches are helpful on labeled datasets only. Recently, several threshold-based software fault prediction approaches have been proposed. However, these approaches do not incorporate the distribution of software metrics for metric threshold derivation; hence, they demonstrate poor performance. To fill this gap, we develop an automated fault prediction approach, namely threshold clustering labeling/threshold clustering labeling plus (TCLP), which does not need a labeled dataset. It can identify the faulty and nonfaulty artifacts on unlabeled datasets by self-learning. Our proposed approach is an extension of the state-of-the-art technique known as CLAMI. Unlike CLAMI, we derive the metrics threshold using logarithmic transformation. Thereafter, we label the instances into binary classes (faulty/nonfaulty) using the metric threshold values. TCLP extends this approach one step further by performing fault prediction using a random forest algorithm. The empirical evaluation of the proposed approach on 28 datasets (with the different number of metrics and granularity) collected from five software groups shows that the proposed unsupervised method obtains significantly better results than those of the state-of-the-art methods. The proposed approach impressively enhances the performance of CLAMI in terms of accuracy, F-measure, and Mathew’s correlation coefficient.

Deep Learning Domain Adaptation for Electro-Mechanical Actuator Fault Diagnosis Under Variable Driving Waveforms

Benefited from the sampling signal from monitoring sensors installed in electro-mechanical actuator (EMA), data-driven methods, such as deep learning models that can mine adaptive features from monitoring signal, have received wide attention in EMA fault diagnosis. A generalized hypothesis accepted by previous efforts on EMA fault diagnosis, is that target domain datasets (i.e., testing monitoring signal) are drawn from the same distribution with source domain datasets (i.e., training monitoring signal). However, except for the different working loads and velocities, various driving signal waveforms also exits in EMA, which further increase the domain discrepancy challenge and decrease the performance of those deep learning models. Therefore, this paper proposes a deep learning domain adaptation method that introduces joint maximum mean discrepancy (JMMD) into convolutional neural network (CNN), which aims to reduce the domain difference between labeled source domain datasets and unlabeled target domain datasets. In order to simulate domain discrepancy phenomenon, the NASA EMA datasets, sampled from acceleration sensors under variable driving waveforms with different working loads and velocities, are considered in this paper. Experiment results indicate that CNN-JMMD method can reduce the distribution difference and improve the diagnosis accuracy in source and target domains. In comparison with several available methods, extensive experimental evaluations on six domain tasks indicate the effectiveness of the proposed method. In addition, better performance has been obtained by selecting the proper trade-off factor between CNN and JMMD.

Memristive devices based hardware for unlabeled data processing

Abstract Unlabeled data processing is of great significance for artificial intelligence (AI), since well-structured labeled data are scarce in a majority of practical applications due to the high cost of human annotation of labeling data. Therefore, automatous analysis of unlabeled datasets is important, and relevant algorithms for processing unlabeled data, such as k-means clustering, restricted Boltzmann machine and locally competitive algorithms etc, play a critical role in the development of AI techniques. Memristive devices offer potential for power and time efficient implementation of unlabeled data processing due to their unique properties in neuromorphic and in-memory computing. This review provides an overview of the design principles and applications of memristive devices for various unlabeled data processing and cognitive AI tasks.

Self-Supervised Graph Neural Networks for Accurate Prediction of Néel Temperature

Antiferromagnetic materials are exciting quantum materials with rich physics and great potential for applications. On the other hand, an accurate and efficient theoretical method is highly demanded for determining critical transition temperatures, Néel temperatures, of antiferromagnetic materials. The powerful graph neural networks (GNNs) that succeed in predicting material properties lose their advantage in predicting magnetic properties due to the small dataset of magnetic materials, while conventional machine learning models heavily depend on the quality of material descriptors. We propose a new strategy to extract high-level material representations by utilizing self-supervised training of GNNs on large-scale unlabeled datasets. According to the dimensional reduction analysis, we find that the learned knowledge about elements and magnetism transfers to the generated atomic vector representations. Compared with popular manually constructed descriptors and crystal graph convolutional neural networks, self-supervised material representations can help us to obtain a more accurate and efficient model for Néel temperatures, and the trained model can successfully predict high Néel temperature antiferromagnetic materials. Our self-supervised GNN may serve as a universal pre-training framework for various material properties.

Towards robust diagnosis of COVID-19 using vision self-attention transformer

The outbreak of COVID-19, since its appearance, has affected about 200 countries and endangered millions of lives. COVID-19 is extremely contagious disease, and it can quickly incapacitate the healthcare systems if infected cases are not handled timely. Several Conventional Neural Networks (CNN) based techniques have been developed to diagnose the COVID-19. These techniques require a large, labelled dataset to train the algorithm fully, but there are not too many labelled datasets. To mitigate this problem and facilitate the diagnosis of COVID-19, we developed a self-attention transformer-based approach having self-attention mechanism using CT slices. The architecture of transformer can exploit the ample unlabelled datasets using pre-training. The paper aims to compare the performances of self-attention transformer-based approach with CNN and Ensemble classifiers for diagnosis of COVID-19 using binary Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) infection and multi-class Hybrid-learning for UnbiaSed predicTion of COVID-19 (HUST-19) CT scan dataset. To perform this comparison, we have tested Deep learning-based classifiers and ensemble classifiers with proposed approach using CT scan images. Proposed approach is more effective in detection of COVID-19 with an accuracy of 99.7% on multi-class HUST-19, whereas 98% on binary class SARS-CoV-2 dataset. Cross corpus evaluation achieves accuracy of 93% by training the model with Hust19 dataset and testing using Brazilian COVID dataset.