Unlabeled Set Research Articles

Generalizable and efficient cross‐domain person re‐identification model using deep metric learning

AbstractMost of the successful person re‐ID models conduct supervised training and need a large number of training data. These models fail to generalise well on unseen unlabelled testing sets. The authors aim to learn a generalisable person re‐identification model. The model uses one labelled source dataset and one unlabelled target dataset during training and generalises well on the target testing set. To this end, after a feature extraction by the ResNext‐50 network, the authors optimise the model by three loss functions. (a) One loss function is designed to learn the features of the target domain by tuning the distances between target images. Therefore, the trained model will be more robust to overcome the intra‐domain variations in the target domain and generalises well on the target testing set. (b) One triplet loss is used which considers both source and target domains and makes the model learn the inter‐domain variations between source and target domain as well as the variations in the target domain. (c) Also, one loss function is for supervised learning on the labelled source domain. Extensive experiments on Market1501 and DukeMTMC re‐ID show that the model achieves a very competitive performance compared with state‐of‐the‐art models and also it requires an acceptable amount of GPU RAM compared to other successful models.

Intelligent Fault Diagnosis With Noisy Labels via Semisupervised Learning on Industrial Time Series

Deep neural networks (DNNs) excel at industrial fault diagnosis. Their performance heavily relies on the quality of human-annotated labels. Due to perception limitations of annotators, industrial time series samples (such as vibration and voltage signals) are frequently mislabeled in several conditions, such as samples with frequency domain feature differences and samples on class borders. Hence, an annotated industrial dataset will inevitably contain noisy labels at a certain level, leading to over-fitting and poor generalization of DNNs. In this work, we introduce an Industrial Noisy Label Semi-Supervised Learning (INL-SSL) fault diagnosis approach, addressing the problem that a certain number of samples in an industrial dataset are mislabeled. The proposed INL-SSL architecture simultaneously trains two DNNs, which cross-train on each other to filter noisy label errors. In particular, a fitted Gaussian mixture model divides time series samples of each DNN flow into an unlabeled set with samples likely to be noisy and a labeled set with samples likely to be clean. Given the labeled and unlabeled data, we proposed a time series MixMatch semi-supervised learning strategy to train the diagnostic model. Ablation study verifies the benefit of the proposed time series augmentation techniques for semi-supervised training. Extensive experiments on a benchmark industrial dataset of rolling element bearings (REB) reveal that the INL-SSL outperforms state-of-the-art approaches. On another self-collected REB dataset, the proposed approach also exceeds other comparison methods under noise ratios from 20% to 90%, validating the model's generalizability.

A Global Review of Publicly Available Datasets Containing Fundus Images: Characteristics, Barriers to Access, Usability, and Generalizability.

This article provides a comprehensive and up-to-date overview of the repositories that contain color fundus images. We analyzed them regarding availability and legality, presented the datasets' characteristics, and identified labeled and unlabeled image sets. This study aimed to complete all publicly available color fundus image datasets to create a central catalog of available color fundus image datasets.

An Accelerator for Semi-Supervised Classification with Granulation Selection

Semi-supervised classification is one of the core methods to deal with incomplete tag information without manual intervention, which has been widely used in various real problems for its excellent performance. However, the existing algorithms need to store all the unlabeled instances and repeatedly use them in the process of iteration. Thus, the large population size may result in slow execution speed and large memory requirements. Many efforts have been devoted to solving this problem, but mainly focused on supervised classification. Now, we propose an approach to decrease the size of the unlabeled instance set for semi-supervised classification algorithms. In this algorithm, we first divide the unlabeled instance set into several subsets with the information granulation mechanism, then sort the divided subsets according to the contribution to the classifier. Following this order, the subsets that take great classification performance are saved. The proposed algorithm is compared with the state-of-the-art algorithms on 12 real datasets, and experiment results show it could get a similar prediction ability but have the lowest instance storage ratio.

ECG signal reconstruction based on facial videos via combined explicit and implicit supervision

Accurate and continuous electrocardiogram (ECG) signal measurements are of paramount importance for the prevention of cardiovascular disease. Existing physiological measurement methods require the placement of sensors on the human body, which proves uncomfortable and inconvenient during long-term monitoring. Recent research has found that heart rates can be extracted from noncontact RGB facial videos by measuring subtle color changes in the skin region. However, designing an accurate feature representation structure for physiological measurements remain a challenge. This study presents a novel end-to-end ECG signal synthesis network using short facial video sequences. We combined explicit and implicit supervision structures, which allows the network to perceive physiological-related features from unlabeled feature sets and adopts a channel and frame aggregation attention mechanism to boost the network acquisition of periodic spatiotemporally correlated features. Finally, the generated features were fed into a physiological estimator consisting of a serially connected “Generative Adversarial” module for ECG signal reconstruction. Experimental results on publicly available datasets and a new dataset collected by us demonstrate that the proposed network outperforms state-of-the-art methods in ECG signal measurements. This is a powerful tool for use in telemedicine and health monitoring.

Automatic facies classification from acoustic image logs using deep neural networks

Borehole image logs greatly facilitate a detailed characterization of rock formations, especially for the highly heterogeneous and anisotropic carbonate rocks. However, interpreting image logs requires massive time and workforce and lacks consistency and repeatability because it relies heavily on a human interpreter’s expertise, experience, and alertness. Thus, we propose to train an end-to-end deep neural network (DNN) for instant and consistent facies classification of carbonate rocks from acoustic image logs and gamma-ray logs. The DNN is modified from the well-known U-Net for image segmentation. The training data are composed of two data sets: (1) manually labeled field data measured by different imaging tools from the geologically complex Brazilian presalt region and (2) noise-free synthetic data. Some short sections of the field data that are challenging for manual labeling due to entangled features and noise or low resolution are left unlabeled for a blind test after training. All labeled data are divided into a training set, a validation set, and a test set to avoid overfitting. We determine that the trained DNN achieves a 77% classification accuracy for the test set and provides reasonable predictions for the challenging unlabeled sets. It is a great achievement given the complexity and variability of the field data. Compared with manual classification, our DNN provides more consistent and higher-resolution predictions in a highly efficient manner and thus dramatically contributes to an automatic image log interpretation.

Label-Free Model Evaluation with Out-of-Distribution Detection

In recent years, label-free model evaluation has been developed to estimate the performance of models on unlabeled test sets. However, we find that existing methods perform poorly in environments with out-of-distribution (OOD) data. To address this issue, we propose a novel automatic model evaluation method using OOD detection to reduce the impact of OOD data on model evaluation. Specifically, we use the representation of datasets to train a neural network for accuracy prediction and employ energy-based OOD detection to exclude OOD data during testing. We conducted experiments on several benchmark datasets with varying amounts of OOD data (SVHN, ISUN, ImageNet, and LSUN) and demonstrated that our method reduces the RMSE compared to existing methods by at least 1.27%. Additionally, we tested our method on transformed datasets and datasets with a high proportion of OOD data, and the results show its robustness.

Instance Segmentation Model Evaluation and Rapid Deployment for Autonomous Driving Using Domain Differences

This paper studies the model evaluation and rapid deployment problems when the model trained by the source domain is deployed to multiple other target domains in autonomous driving application scenarios. To address the problem that the existing methods for evaluating model accuracy require complete labeled test sets, this paper proposes an instance segmentation model evaluation method based on domain differences, which can give the prediction accuracy of the model on unlabeled test sets. Furthermore, to address the rapid model deployment, this paper proposes a calculation method for the minimum number of supplements based on domain differences. The method selects the fewest images from the target domain scene and realizes the rapid deployment of the model through further annotation and training. Finally, experiments are conducted on the public dataset Cityscapes. The model evaluation experiments show that the root mean square error (RMSE) of the proposed method on Cityscapes is about 4% smaller (from 6% to 2%) than that of other existing model evaluation methods, and it can be as small as 2.35%. Rapid deployment experiments show that the RMSE of the proposed method on Cityscapes is about 8%.

Optimal unlabeled set partitioning with application to risk-based quarantine policies

We consider the problem of partitioning a set of items into unlabeled subsets so as to optimize an additive objective, i.e., the objective function value of a partition is equal to the sum of the contribution of each subset. Under an arbitrary objective function, this family of problems is known to be an -complete combinatorial optimization problem. We study this problem under a broad family of objective functions characterized by elementary symmetric polynomials, which are “building blocks” to symmetric functions. By analyzing a continuous relaxation of the problem, we identify conditions that enable the use of a reformulation technique in which the set partitioning problem is cast as a more tractable network flow problem solvable in polynomial-time. We show that a number of results from the literature arise as special cases of our proposed framework, highlighting its generality. We demonstrate the usefulness of the developed methodology through a novel and timely application of quarantining heterogeneous populations in an optimal manner. Our case study on real COVID-19 data reveals significant benefits over conventional measures in terms of both spread mitigation and economic impact, underscoring the importance of data-driven policies.

Subsequence and distant supervision based active learning for relation extraction of Chinese medical texts

In recent years, relation extraction on unstructured texts has become an important task in medical research. However, relation extraction requires a large amount of labeled corpus, manually annotating sequences is time consuming and expensive. Therefore, efficient and economical methods for annotating sequences are required to ensure the performance of relational extraction. This paper proposes a method of subsequence and distant supervision based active learning. The method is annotated by selecting information-rich subsequences as a sampling unit instead of the full sentences in traditional active learning. Additionally, the method saves the labeled subsequence texts and their corresponding labels in a dictionary which is continuously updated and maintained, and pre-labels the unlabeled set through text matching based on the idea of distant supervision. Finally, the method combines a Chinese-RoBERTa-CRF model for relation extraction in Chinese medical texts. Experimental results test on the CMeIE dataset achieves the best performance compared to existing methods. And the best F1 value obtained between different sampling strategies is 55.96%.

Pyramidal Dilation Attention Convolutional Network With Active and Self-Paced Learning for Hyperspectral Image Classification

In recent years, deep neural networks have been widely used for hyperspectral image (HSI) classification and have shown excellent performance using numerous labeled samples. The acquisition of HSI labels is usually based on the field investigation, which is expensive and time consuming. Hence, the available labels are usually limited, which affects the efficiency of deep HSI classification methods. To improve the classification performance while reducing the labeling cost, this article proposes a semisupervised deep learning (DL) method for HSI classification, named pyramidal dilation attention convolutional network with active and self-paced learning (PDAC-ASPL), which integrates active learning (AL), self-paced learning (SPL), and DL into a unified framework. First, a densely connected pyramidal dilation attention convolutional network is trained with a limited number of labeled samples. Then, the most informative samples from the unlabeled set are selected by AL and queried real labels, and the highest confidence samples with corresponding pseudo labels are extracted by SPL. Finally, the samples from AL and SPL are added to the training set to retrain the network. Compared with some DL- and AL-based HSI classification methods, our PDAC-ASPL achieves better performance on four HSI datasets.

Multiple Instance Differentiation Learning for Active Object Detection.

Despite the substantial progress of active learning for image recognition, there lacks a systematic investigation of instance-level active learning for object detection. In this paper, we propose to unify instance uncertainty calculation with image uncertainty estimation for informative image selection, creating a multiple instance differentiation learning (MIDL) method for instance-level active learning. MIDL consists of a classifier prediction differentiation module and a multiple instance differentiation module. The former leverages two adversarial instance classifiers trained on the labeled and unlabeled sets to estimate instance uncertainty of the unlabeled set. The latter treats unlabeled images as instance bags and re-estimates image-instance uncertainty using the instance classification model in a multiple instance learning fashion. Through weighting the instance uncertainty using instance class probability and instance objectness probability under the total probability formula, MIDL unifies the image uncertainty with instance uncertainty in the Bayesian theory framework. Extensive experiments validate that MIDL sets a solid baseline for instance-level active learning. On commonly used object detection datasets, it outperforms other state-of-the-art methods by significant margins, particularly when the labeled sets are small. The code is available at https://github.com/WanFang13/MIDL.

Remote Sensing Image Scene Classification with Self-Supervised Learning Based on Partially Unlabeled Datasets

In recent years, supervised learning, represented by deep learning, has shown good performance in remote sensing image scene classification with its powerful feature learning ability. However, this method requires large-scale and high-quality handcrafted labeled datasets, which leads to a high cost of obtaining annotated samples. Self-supervised learning can alleviate this problem by using unlabeled data to learn the image’s feature representation and then migrate to the downstream task. In this study, we use an encoder–decoder structure to construct a self-supervised learning architecture. In the encoding stage, the image mask is used to discard some of the image patches randomly, and the image’s feature representation can be learned from the remaining image patches. In the decoding stage, the lightweight decoder is used to recover the pixels of the original image patches according to the features learned in the encoding stage. We constructed a large-scale unlabeled training set using several public scene classification datasets and Gaofen-2 satellite data to train the self-supervised learning model. In the downstream task, we use the encoder structure with the masked image patches that have been removed as the backbone network of the scene classification task. Then, we fine-tune the pre-trained weights of self-supervised learning in the encoding stage on two open datasets with complex scene categories. The datasets include NWPU-RESISC45 and AID. Compared with other mainstream supervised learning methods and self-supervised learning methods, our proposed method has better performance than the most state-of-the-art methods in the task of remote sensing image scene classification.

Semi-supervised learning with dropouts

Graph Laplacian manifold regularization exploits the intrinsic geometrical properties of the underlying graph over both labeled and unlabeled data to approximate an optimal smooth prediction function. The regularization works on the assumption that abundant unlabeled data add information to the model, which increases accuracy and generalization. However, a large number of unlabeled data points with few labeled data points results in function degeneration which restricts generalization. Higher order regularization through iterated Laplacian semi-norm resolves this problem, but it is not computationally intensive and not scalable. In this paper, we propose dropouts regularization by appropriately balancing the supervised loss and intrinsic regularization terms to enforce smoothness. This is done by increasing the labeled data points by assigning temporary soft-labels to a subset of unlabeled set iteratively. This is followed by the determination of the final labels. Localized Procrustes analysis is, then, used to increase the accuracy of label assignment. Experiments on both synthetic and real-world dataset show that the proposed model outperforms the existing graph Laplacian and its variants by a significant margin.

Robust image hashing for content identification through contrastive self-supervised learning

Content identification systems are an essential technology for many applications. These systems identify query multimedia items using a database of known identities. A hash-based system uses a perceptual hashing function that generates a hash value invariant against a set of expected manipulations in an image, later compared to perform identification. Usually, this set of manipulations is well-known, and the researcher creates the perceptual hashing function that best adapts to the set. However, a new manipulation may break the hashing function, requiring to create a new one, which may be costly and time-consuming. Therefore, we propose to let the hashing function learn an invariant feature space automatically. For this, we exploit the recent advances in self-supervised learning, where a model uses unlabeled data to generate a feature representation by solving a metric learning-based pretext task that enforces the robust image hashing properties for content identification systems. To achieve model transferability on unseen data, our pretext task enforces the feature vector invariance against the manipulation set, and through random sampling on the unlabeled training set, we present the model a wide variety of perceptual information to work on. As exhaustive experimentation shows, this method achieves excellent robustness against a comprehensive set of manipulations, even difficult ones such as horizontal flip and rotation, with excellent identification performance. Also, the trained model is highly discriminative against the presence of near-duplicate images. Furthermore, this method does not need re-training or fine-tuning on a new dataset to achieve the observed performance, indicating an excellent generalization capacity.

It is not all about salary: a discrete-choice experiment to determine community health workers’ motivation for work in Nigeria

IntroductionCommunity health workers (CHWs) constitute the majority of primary healthcare (PHC) workers in Nigeria, yet little is understood about their motivations or the most effective interventions to meet their needs...

SelfNet: A semi-supervised local Fisher discriminant network for few-shot learning

Few-shot learning, employing small-scale labeled samples to recognize new objects, has received substantial research interest. The prototypical network (ProtoNet) is a simple yet effective meta-learning method to solve this problem. In the few-shot scenario, however, the scarcity of data usually has a negative impact on the representational ability of prototypes. In this paper, a unique semi-supervised few-shot learning architecture, referred to as Semi-supervised local Fisher discriminant network (SelfNet), which integrates few-shot learning with subspace learning, is proposed. Using the union of the support set and the additional unlabeled set, a feature projection module is constructed to achieve the subspace projection. Additionally, a pseudo-labeling strategy, which adds the unlabeled samples with high prediction confidence to the support set, is employed to refine the original prototypes. Experimental results on two few-shot classification benchmarks demonstrate that SelfNet can achieve superior performance to the state-of-the-arts, indicating the benefits of utilizing unlabeled samples for feature projection.

LongReMix: Robust learning with high confidence samples in a noisy label environment

State-of-the-art noisy-label learning algorithms rely on an unsupervised learning to classify training samples as clean or noisy, followed by a semi-supervised learning (SSL) that minimises the empirical vicinal risk using a labelled set formed by samples classified as clean, and an unlabelled set with samples classified as noisy. The classification accuracy of such noisy-label learning methods depends on the precision of the unsupervised classification of clean and noisy samples, and the robustness of SSL to small clean sets. We address these points with a new noisy-label training algorithm, called LongReMix, which improves the precision of the unsupervised classification of clean and noisy samples and the robustness of SSL to small clean sets with a two-stage learning process. The stage one of LongReMix finds a small but precise high-confidence clean set, and stage two augments this high-confidence clean set with new clean samples and oversamples the clean data to increase the robustness of SSL to small clean sets. We test LongReMix on CIFAR-10 and CIFAR-100 with introduced synthetic noisy labels, and the real-world noisy-label benchmarks CNWL (Red Mini-ImageNet), WebVision, Clothing1M, and Food101-N. The results show that our LongReMix produces significantly better classification accuracy than competing approaches, particularly in high noise rate problems. Furthermore, our approach achieves state-of-the-art performance in most datasets. The code is available at https://github.com/filipe-research/LongReMix.

Boosting semi-supervised face recognition with raw faces

Deep facial recognition benefits significantly from large-scale training data; however, the bottleneck of high labeling costs persists. Therefore, to reduce the labeling costs, it is desirable to train a model using limited labeled data and abundant unlabeled data ( i.e. , semi-supervised learning). However, existing semi-supervised learning methods present two primary challenges: (1) The possibility of identity overlaps between the unlabeled and labeled data. These overlaps can affect the correctness of pseudo-labels of the unlabeled set. (2) Different pseudo-labels generated by the clustering algorithm may belong to the same individual ( i.e. , over-decomposition problem). Thus, in this study, instead of experimenting with non-overlapping conditions, we apply smooth labels to exploit the potential of those samples that are similar to the identities in the labeled set. For samples that are not similar to the labeled set, we introduce a dual clustering strategy to remedy the over-decomposition problem caused by single clustering. With the upgraded semi-supervised framework, we recycle the discarded samples during purification of MS-Celeb-1 M (MS1M) to further scale up the training set, which offers a considerable performance boost of 94.39% on the IJB-C dataset. • Pioneering use of raw unlabeled data. • Utilizing of overlaps by relaxing the intra-class constraint with label smoothing. • Dual clustering strategy to address the over decomposition problem caused by a single clustering.

Classical Planning in Deep Latent Space

Current domain-independent, classical planners require symbolic models of the problem domain and instance as input, resulting in a knowledge acquisition bottleneck. Meanwhile, although deep learning has achieved significant success in many fields, the knowledge is encoded in a subsymbolic representation which is incompatible with symbolic systems such as planners. We propose Latplan, an unsupervised architecture combining deep learning and classical planning. Given only an unlabeled set of image pairs showing a subset of transitions allowed in the environment (training inputs), Latplan learns a complete propositional PDDL action model of the environment. Later, when a pair of images representing the initial and the goal states (planning inputs) is given, Latplan finds a plan to the goal state in a symbolic latent space and returns a visualized plan execution. We evaluate Latplan using image-based versions of 6 planning domains: 8-puzzle, 15-Puzzle, Blocksworld, Sokoban and Two variations of LightsOut.