Manual Labeling Research Articles

Cloud detection sample generation algorithm for nighttime satellite imagery based on daytime data and machine learning application.

Highly accurate nighttime cloud detection in satellite imagery is challenging due to the absence of visible to near-infrared (0.38-3μm, VNI) data, which is critical for distinguishing clouds from other ground features. Fortunately, Machine learning (ML) techniques can more effectively leverage the limited wavelength information and show high-accuracy cloud detection based on vast sample volume. However, accurately distinguishing cloud pixels solely through thermal infrared bands (8-14μm, TIR) is challenging, and acquiring numerous, high-quality and representative samples of nighttime images for ML proves to be unattainable in practice. Given the thermal infrared radiation transmission process and the fact that daytime and nighttime have the same source of radiance in TIR, we propose a sample generation idea that uses daytime images to provide samples of nighttime cloud detection, which is different from the traditional sample construction methods (e.g., manually label, simulation and transfer learning method), and can obtain samples effectively. Based on this idea, nighttime cloud detection experiments were carried out for MODIS, GF-5 (02) and Himawari-8 satellites, respectively. The results were validated by the Lidar cloud product and manual labels and show that our nighttime cloud detection result has higher accuracy than MYD35 (78.17%) and the ML model trained by nighttime manual labels (75.86%). The accuracy of the three sensors is 82.19%, 88.71%, and 79.34%, respectively. Moreover, we validated and discussed the performance of our algorithm on various surface types (vegetation, urban, barren and water). The results revealed that the accuracy of the three sensors over barren was found to be poor and varied with surface types but overall high. Our study can provide a novel perspective on nighttime cloud detection of muti-spectral satellite imagery.

Fracture mode classification of UHPC based on deep learning and wavelet time–frequency spectrum derived from acoustic emission waveform

Fracture mode categorization is essential in identifying the failure stage, evaluating the damage characteristics, and providing references for structural maintenance. Acoustic emission (AE), as a passive nondestructive technique, presents great potential in damage characterization of cementitious materials and structures. In this study, a convolutional neural network (CNN) and gated recurrent unit (GRU) integrated model for tensile, shear, and mixed cracking mode classification of ultra-high-performance concrete (UHPC) was proposed and verified based on AE waveform features. First, four-point bending and direct shear tests for UHPC specimens were carried out, respectively, to generate AE waveform signals corresponding to different cracking modes (tensile, shear, and mixed ones). Subsequently, the collected one-dimensional AE signals were converted into two-dimensional time–frequency spectrum by continuous wavelet transform. And an AE dataset was constructed after manual labeling based on the corresponding relationships between cracking modes and wavelet time–frequency spectra. Then, an automatic CNN-GRU automatic classification model was established using wavelet time–frequency spectrum images as input variables, where the CNN extracts spatial image features, and the GRU screens time dependencies and realizes final classification. The proposed CNN-GRU model achieves an accuracy of 96.17% for cracking mode classification of UHPC, which is superior both in computational accuracy and efficiency to lightweight CNN and AlexNet. Classification outcomes for UHPC specimen under four-point bending and direct shear tests revealed the proportion of tensile cracks progressively decreased with the increase of damage, while shear cracks exhibited the opposite trend. The percentage of mixed mode cracks maintained relatively stable throughout the failure stages. The model proposed in this study presents favorable performance to reveal the underlying damage evolution mechanism during UHPC fracture, which can also provide references for the cracking mode classification of other cementitious materials and structures.

Pseudo-Labeling and Time-Series Data Analysis Model for Device Status Diagnostics in Smart Agriculture

This study proposes an automated data-labeling model that combines a pseudo-labeling algorithm with waveform segmentation based on Long Short-Term Memory (LSTM) to effectively label time-series data in smart agriculture. This model aims to address the inefficiency of manual labeling for large-scale data generated by agricultural systems, enhancing the performance and scalability of predictive models. Our proposed method leverages key features of time-series data to automatically generate labels for new data, thereby improving model accuracy and streamlining data processing. By automating the labeling process, we reduce dependence on manual labeling, which is often labor-intensive and prone to errors in large datasets. This approach enables the efficient preparation of labeled data for applications such as anomaly detection, pattern recognition, and predictive modeling in smart agriculture. Experimental results demonstrate that the automated labeling model achieves 89% accuracy in agricultural environments and reduces data processing time by 30%. Future research will focus on extending the model’s applicability to diverse agricultural settings, enhancing generalization performance, and improving real-time processing capabilities, thereby advancing intelligent and sustainable smart agriculture systems.

Validation of Radiosynthesis and First in-Human Dosimetry of 68Ga-NOTA-UBI-29-41: A Proof of Concept Study.

Background: Antimicrobial peptides (AMPs) such as UBI-29-41 offer a distinctive approach for precise detection due to their unique interactions with bacteria and makes them promising candidates for specific and selective imaging. The study was aimed to corroborate the in-house manual synthesis of 68Ga-NOTA-UBI-29-41, evaluate its uptake in patients with suspected infection, and estimate of patient-specific dosimetry to ensure optimal clinical application. Materials and Methods: 68Ga-NOTA-UBI-29-41 was synthesized by using a variable amount of UBI-29-41 (60-90 μg) to 555 MBq of Ga-68 in 0.05 M Hydrochloric acid (HCl) and heating the reaction sample for 12 min at 90°C at pH: 3.5-4 to obtain the radiopeptide with high yield and high radiochemical purity (RCP). 68Ga-NOTA-UBI-29-41 positron emission tomography/Computed tomography (CT) scans at variable timepoints were done to evaluate its biodistribution and maximum uptake time. Furthermore, patient-specific dosimetric estimation was done using the HERMES software. Results: A total of 5 μg/37 MBq (5 μg/mCi) of NOTA-UBI-29-41 for 12 min at 90°C were the optimal parameters to obtain 88%-90% of yield and 98%-99 % of RCP. 68Ga-NOTA-UBI-29-41 showed expeditious blood clearance and high renal excretion. The optimal time for imaging of infection with 68Ga-NOTA-UBI-29-41 was found to be at 60 min postinjection (n = 8). The critical organ was the urinary bladder, receiving an average dose of 138.02 ± 45.92 µSv/MBq, followed by 53.81 ± 13.72 µSv/MBq for kidneys with a mean effective dose of 1.52 ± 0.64 mSv. Conclusion: The protocol for in-house manual labeling of 68Ga-NOTA-UBI-29-41 was reproducible, providing high yield and RCP. 68Ga-NOTA-UBI-29-41 administration was found to be safe and nontoxic. The favorable biodistribution and the first-in-human patient-specific dosimetry ensure optimal clinical application.

Quantitative modeling of lenticulostriate arteries on 7-T TOF-MRA for cerebral small vessel disease.

We developed a framework for segmenting and modeling lenticulostriate arteries (LSAs) on 7-T time-of-flight magnetic resonance angiography and tested its performance on cerebral autosomal dominant arteriopathy with subcortical infarcts and leukoencephalopathy (CADASIL) patients and controls. We prospectively included 29 CADASIL patients and 21 controls. The framework includes a small-patch convolutional neural network (SP-CNN) for fine segmentation, a random forest for modeling LSAs, and a screening model for removing wrong branches. The segmentation performance of our SP-CNN was compared to competitive networks. External validation with different resolution was performed on ten patients with aneurysms. Dice similarity coefficient (DSC) and Hausdorff distance (HD) between each network and manual segmentation were calculated. The modeling results of the centerlines, diameters, and lengths of LSAs were compared against manual labeling by four neurologists. The SP-CNN achieved higher DSC (92.741 ± 2.789, mean ± standard deviation) and lower HD (0.610 ± 0.141 mm) in the segmentation of LSAs. It also outperformed competitive networks in the external validation (DSC 82.6 ± 5.5, HD 0.829 ± 0.143 mm). The framework versus manual difference was lower than the manual inter-observer difference for the vessel length of primary branches (median -0.040 mm, interquartile range -0.209 to 0.059 mm) and secondary branches (0.202 mm, 0.016-0.537 mm), as well as for the offset of centerlines of primary branches (0.071 mm, 0.065-0.078 mm) and secondary branches (0.072, 0.064-0.080 mm), with p < 0.001 for all comparisons. Our framework for LSAs modeling/quantification demonstrated high reliability and accuracy when compared to manual labeling. NCT05902039 ( https://clinicaltrials.gov/study/NCT05902039?cond=NCT05902039 ). The proposed automatic segmentation and modeling framework offers precise quantification of the morphological parameters of lenticulostriate arteries. This innovative technology streamlines diagnosis and research of cerebral small vessel disease, eliminating the burden of manual labeling, facilitating cohort studies and clinical diagnosis. The morphology of LSAs is important in the diagnosis of CSVD but difficult to quantify. The proposed algorithm achieved the performance equivalent to manual labeling by neurologists. Our method can provide standardized quantitative results, reducing radiologists' workload in cohort studies.

Manual data labeling, radiology, and artificial intelligence: It is a dirty job, but someone has to do it

Manual data labeling, radiology, and artificial intelligence: It is a dirty job, but someone has to do it

Polytope Fraud Theory

Polytope Fraud Theory (PFT) extends the existing triangle and diamond theories of accounting fraud with ten abnormal financial practice alarms that a fraudulent firm might trigger. These warning signals are identified through evaluation of the shorting behavior of sophisticated activist short sellers, which are used to train several supervised machine learning methods in detecting financial statement fraud using published accounting data. Our contributions include a systematic manual collection and labeling of companies that are shorted by professional activist short sellers. We also combine well-known asset pricing factors with accounting red flags in financial features selections. Using 80 % of the data for training and the remaining 20 % for out-of-sample test and performance assessment, we find that the best method is XGBoost, with a Recall of 72 % and F1-score of 82 %. Other methods have relatively lower performance, demonstrating the robustness of our results. This shows that the sophisticated activist short sellers, from whom the algorithms are learning, have excellent accounting insights, tremendous forensic analytical knowledge, and sharp business acumen. Our feature importance analysis indicates that potential short-selling targets share many similar financial characteristics, such as bankruptcy or financial distress risk, clustering in some industries, inconsistency of profitability, high accrual, and unreasonable business operations. Our results imply the possible automation of advanced financial statement analysis, which can both improve auditing processes and effectively enhance investment performance. Finally, we propose the Unified Investor Protection Framework, summarizing and categorizing investor-protection related theories from the macro-level to the micro-level.

Employing the Artificial Intelligence Object Detection Tool YOLOv8 for Real-Time Pain Detection: A Feasibility Study.

Effective pain management is crucial for patient care, impacting comfort, recovery, and overall well-being. Traditional subjective pain assessment methods can be challenging, particularly in specific patient populations. This research explores an alternative approach using computer vision (CV) to detect pain through facial expressions. The study implements the YOLOv8 real-time object detection model to analyze facial expressions indicative of pain. Given four pain datasets, a dataset of pain-expressing faces was compiled, and each image was carefully labeled based on the presence of pain-associated Action Units (AUs). The labeling distinguished between two classes: pain and no pain. The pain category included specific AUs (AU4, AU6, AU7, AU9, AU10, and AU43) following the Prkachin and Solomon Pain Intensity (PSPI) scoring method. Images showing these AUs with a PSPI score above 2 were labeled as expressing pain. The manual labeling process utilized an open-source tool, makesense.ai, to ensure precise annotation. The dataset was then split into training and testing subsets, each containing a mix of pain and no-pain images. The YOLOv8 model underwent iterative training over 10 epochs. The model's performance was validated using precision, recall, and mean Average Precision (mAP) metrics, and F1 score. When considering all classes collectively, our model attained a mAP of 0.893 at a threshold of 0.5. The precision for "pain" and "nopain" detection was 0.868 and 0.919, respectively. F1 scores for the classes "pain", "nopain", and "all classes" reached a peak value of 0.80. Finally, the model was tested on the Delaware dataset and in a real-world scenario. Despite limitations, this study highlights the promise of using real-time computer vision models for pain detection, with potential applications in clinical settings. Future research will focus on evaluating the model's generalizability across diverse clinical scenarios and its integration into clinical workflows to improve patient care.

Superpixel-based principal feature clustering annotation method for dual-phase microstructure segmentation

Metallographic analysis is one of the most commonly used techniques by materials scientists for studying metal materials. The deep learning methods, which have been widely applied in metallographic images analysis, demonstrate excellent performance in this task. However, the optimization of deep learning models often relies on a substantial amount of accurately labeled samples for effective supervision. To address this issue, this paper proposes a novel automatic annotation method based on brightness and spatial distribution which is suitable for deep learning-based segmentation of optical dual-phase metallographic microstructure. The proposed automatic annotated method includes the superpixel segmentation, principal feature extraction, and clustering algorithm therefore it is referred as the superpixel-based principal feature clustering annotation (SPFCA) method. SPFCA employs discriminative criteria similar to those used by metallurgists to differentiate between metallographic structures. Furthermore, it can mitigate the occasional errors inherent in manual annotation, leading to improved performance compared to models trained with expert annotations. Experimental validation was conducted using four self-built datasets with different image qualities to test the performance of models from different perspective. Initially, hyperparameter optimization for the SPFCA method tailored to our dataset was performed. Subsequently, SPFCA was utilized to guide the optimization of the convolutional neural network employed for segmentation. The results demonstrate that the segmentation model optimized with SPFCA guidance achieved an F1 score of 0.9226 in the single dataset without the need for manual labeling, surpassing the segmentation models optimized with expert annotations.

A novel automated labelling algorithm for deep learning-based built-up areas extraction using nighttime lighting data

The use of remote sensing imagery and cutting-edge deep learning techniques can produce impressive results when it comes to built-up areas extraction (BUAE). However, reducing the manual labelling set production process while ensuring high accuracy is currently the main research topic. This study pioneers the exploitation of nighttime lighting data (NLD) for automatically generating deep learning label sets, assessing the feasibility, and identifying limitations of using varied intensity ranges of lighting data directly for this purpose. We provide a novel method for generating fine-grained labels through an optimisation technique that eliminates the necessity for human involvement. This approach employs deep learning segmentation algorithms and has been tested in eight cities across seven countries. The results indicate that segmentation performs well in most cities, with the combination of iso clustering and NLD allowing for more precise extraction of urban building districts. The overall accuracy exceeds 90% in most cities. The results based on manual and historical data (around 0.7) as labels are significantly lower than those based on NLD. At the same time, the segmentation effect of deep learning has a more significant advantage over traditional machine learning classification algorithms (around 0.8). DeeplabV3 and U-Net exhibit different strengths in segmentation and extraction: DeeplabV3 has a stronger ability to eliminate errors, while U-Net retains the capability to handle less labelled information, making them mutually advantageous depending on the specific requirements of the task. It proposes a Strategy to automatically extract built-up areas with minimal human involvement.

Research on in-vivo electron paramagnetic resonance spectrum classification and radiation dose prediction based on machine learning

The in-vivo electron paramagnetic resonance (EPR) method can be used for on-site, rapid, and non-invasive detection of radiation dose to casualties after nuclear and radiation emergencies. For in-vivo EPR spectrum analysis, manual labeling of peaks and calculation of signal intensity are often used, which have problems such as large workload and interference by subjective factors. In this study, a method for automatic classification and identification of in-vivo EPR spectra was established using support vector machine (SVM) technology, which can in-batch and automatically identify and screen out invalid spectra due to vibration and dental surface water interference during in-vivo EPR measurements. In this study, a spectrum analysis method based on genetic algorithm optimization neural network (GA-BPNN) was established, which can automatically identify the radiation-induced signals in in-vivo EPR spectra and predict the radiation doses received by the injured. The experimental results showed that the SVM and GA-BPNN spectrum processing methods established in this study could effectively accomplish the automatic spectra classification and radiation dose prediction, and could meet the needs of dose assessment in nuclear emergency. This study explored the application of machine learning methods in EPR spectrum processing, improved the intelligence level of EPR spectrum processing, and would help to enhance the efficiency of mass EPR spectra processing.

Automated tree crown labeling with 3D radiative transfer modelling achieves human comparable performances for tree segmentation in semi-arid landscapes

Mapping tree crowns in arid or semi-arid areas, which cover around one-third of the Earth’s land surface, is a key methodology towards sustainable management of trees. Recent advances in deep learning have shown promising results for tree crown segmentation. However, a large amount of manually labeled data is still required. We here propose a novel method to delineate tree crowns from high resolution satellite imagery using deep learning trained with automatically generated labels from 3D radiative transfer modeling, intending to reduce human annotation significantly. The methodological steps consist of 1) simulating images with a 3D radiative transfer model, 2) image style transfer learning based on generative adversarial network (GAN) and 3) tree crown segmentation using U-net segmentation model. The delineation performances of the proposed method have been evaluated on a manually annotated dataset consisting of more than 40,000 tree crowns. Our approach, which relies solely on synthetic images, demonstrates high segmentation accuracy, with an F1 score exceeding 0.77 and an Intersection over Union (IoU) above 0.64. Particularly, it achieves impressive accuracy in extracting crown areas (r2 greater than 0.87) and crown densities (r2 greater than 0.72), comparable to that of a trained dataset with human annotations only. In this study, we demonstrated that the integration of a 3D radiative transfer model and GANs for automatically generating training labels can achieve performances comparable to human labeling, and can significantly reduce the time needed for manual labeling in remote sensing segmentation applications.

Skin feature point tracking using deep feature encodings

AbstractFacial feature tracking is a key component of imaging ballistocardiography (BCG) where accurate quantification of the displacement of facial keypoints is needed for good heart rate estimation. Skin feature tracking enables video-based quantification of motor degradation in Parkinson’s disease. While traditional computer vision algorithms like Scale Invariant Feature Transform (SIFT), Speeded-Up Robust Features (SURF), and Lucas-Kanade method (LK) have been benchmarks due to their efficiency and accuracy, they often struggle with challenges like affine transformations and changes in illumination. In response, we propose a pipeline for feature tracking, that applies a convolutional stacked autoencoder to identify the most similar crop in an image to a reference crop containing the feature of interest. The autoencoder learns to represent image crops into deep feature encodings specific to the object category it is trained upon. We train the autoencoder on facial images and validate its ability to track skin features in general using manually labelled face and hand videos of small and large motion recorded in our lab. Our evaluation protocol is comprehensive, including quantification of errors in human annotation. The tracking errors of distinctive skin features (moles) are so small that we cannot exclude the fact that they stem from the manual labelling based on a $$\chi ^2$$ χ 2 -test. With a mean error of 0.6–3.3 pixels, our method outperformed the other methods in all but one scenario. More importantly, our method was the only one that did not diverge. We also compare our method with the latest state-of-the-art transformer for feature matching by Google—Omnimotion. Our results indicate that our method is superior at tracking different skin features under large motion conditions and that it creates better feature descriptors for tracking, matching, and image registration compared to both traditional algorithms and the latest Omnimotion.

Mapping landforms of a hilly landscape using machine learning and high-resolution LiDAR topographic data

Landform maps are important tools in assessment of soil- and eco-hydrogeomorphic processes and hazards, hydrological modeling, and natural resources and land management. Traditional techniques of mapping landforms based on field surveys or from aerial photographs can be time and labor intensive, highlighting the importance of remote sensing products based automatic or semi-automatic approaches. In addition, the time-intensive manual labeling can also be subjective rather than an objective identification of the landform. Here we implemented such an objective approach applying a random forest machine learning algorithm to a set of observed landform data and 1m horizontal resolution bare-earth digital elevation model (DEM) developed from airborne light detection and ranging (LiDAR) data to rapidly map various landforms of a hilly landscape. The landform classification includes upland plateaus, ridges, convex slopes, planar slopes, concave slopes, stream channels, and valley bottoms, across a 400 km2 hilly landscape of the Ozark Mountains in northeastern Oklahoma. We used 4200 landform observations (600 per landform) and eight topographic indices derived from 2 m, 5 m and 10 m resolution LiDAR DEM in random forest algorithm to develop 2 m, 5 m and 10 m resolution landform models. We test the effectiveness of DEM resolution in mapping landforms via comparison of observed landforms with modeled landforms. Results showed that the approach mapped ∼84% of observed landforms when covariates were at 2 m resolution to ∼89% when they were at 10 m resolution. However, predicted maps showed that the 2 m resolution covariates performed better at capturing accurate landform boundaries and details of small-sized landforms such as stream channels and ridges. The approach presented here significantly reduces the time required for mapping landforms compared to traditional techniques using aerial imagery and field observations and allows incorporation of a wide variety of covariates. The landform map developed using this approach has several potential applications. It could be utilized in hydrological modeling, natural resource management, and characterizing soil-geomorphic processes and hazards in a hilly landscape.

Debiased hybrid contrastive learning with hard negative mining for unsupervised person re-identification

The goal of unsupervised person re-identification is to retrieve a specific person across several non-overlapping cameras without the aid of manual labeling information. In recent times, contrastive learning has found extensive application in undertaking the complexities of unsupervised person Re-ID. Nevertheless, prevailing approaches often ignore the bias in negative proxy sampling and the significance of hard negatives in contrastive learning. These limitations have constrained the performance of existing methods. To solve these issues, we introduce a Debiased Hybrid Contrastive Learning with Hard Negative Mining (DHCL-HNM) approach. Particularly, the proposed approach employs an instance-level memory bank to save the class prototypes for all training images. In each training epoch, the memory bank undergoes clustering, dividing the dataset into un-clustered outliers and clustered images with pseudo labels. Then, the debiasing of negative proxies and the hard negative mining are integrated into a hybrid contrastive learning process to enhance the intra-class similarity and the instance discrimination. The debiasing operation is realized during the sampling of negative proxies to reduce the negative effects of false negatives. In the meantime, the hard negative mining can guide the Re-ID model to concentrate on the hard negatives by reweighting negative proxies based on their similarities to the anchor sample. The efficiency of the proposed method in the realm of unsupervised person Re-ID is demonstrated through comprehensive experiment outcomes conducted on several datasets.

Learning Terrain Traversability for a Mobile Robot based on Information Fusion

In this work, we propose an approach to determine terrain traversability for a car-like robot. Our approach has two main modules: a neural network classifier that makes use of sensors' readings to assign traversability levels to control inputs of the robot, and a second neural network that, based on the outputs of the first network, mimics the control selection performed by a human driver. The approach incorporates sensor fusion from a variety of sources to enhance the traversability estimation, and it is trained employing a semi-supervised learning scheme with examples resulting from the interaction of the car with the environment. This semi-supervised scheme avoids exhausting manual labeling and is built on the premise that there is a correlation between the terrain traversability and the required and observed behaviors of the vehicle. The method is validated with data obtained from a physical electric car.

Adaptive group sample with central momentum contrast loss for unsupervised individual identification of cows in changeable conditions

Individual identification is the basis for accurate breeding and management of dairy cows. Most of the studies on individual identification of dairy cows focused on closed data sets. Meanwhile, the training process requires lots of manual labeling data as support. Aiming at solving the above shortcomings, an unsupervised cow individual identification model named AC-Cowid (Adaptive Group Sample-Central Momentum Contrastive loss-Cowid) was proposed. The innovation of the model lied in the novel sampler and loss function. Adaptive Group Sampler made up for the sampling imbalance problem in the current samplers. The sampler would self-update with model optimization. The central constraint term in Central Momentum Contrastive loss made the updated features more inclined to the cluster center, effectively solved the discomfort of Contrastive loss to cow image data. Feature similarity matching was adopted to meet the challenge of data conversion from a closed set to an open and changeable set. The unsupervised training reduced the workload of manual data labeling. The CMC-1 and mAP index of the final model were 99.50 % and 77.90 %, respectively. Ablation and hyper-parameter experiments proved the effectiveness of the proposed method. The comparison experiment tested eight unsupervised methods and five supervised algorithms in the state-of-the-art. The mAP and CMC-1 of the proposed method was the best among the comparison algorithms. Above all, the proposed method was 4.60 % and 5.38 % higher on mAP and CMC-1 compared with the best algorithm in supervised learning. This study could provide technical support for individual identification of cows in large-scale cattle farms.

Improving acoustic species identification using data augmentation within a deep learning framework

Convolutional neural networks (CNNs) are effective tools for acoustic classification tasks such as species identification. Large datasets of labelled recordings are required to develop CNN classifiers which can be difficult to obtain, particularly if species are rare or vocalise infrequently. Additionally, data often requires manual labelling which can be time consuming requiring expert analysis. Artificially generating data using augmentation can address these challenges, however the impact of data augmentation on CNN performance is poorly understood and often omitted in bioacoustic studies. Here, we empirically test the impact of CNN architecture and 20 data augmentation methods on classifier performance. We use acoustic identification of 18 small mammal species as a case study of a species group that can be effectively surveyed by acoustic monitoring, but recordings for training data are scarce and difficult to collect. Networks that achieved the highest accuracy across all sample sizes was a 10-layer CNN (96.43 %) and a pre-trained ResNet50 model (96.37 %). Overall, all augmentation effects improved ResNet50 model performance and 17 effects improved Conv10 performance, increasing relative change in accuracy (RCA) by 0.021–0.641. Three augmentation effects negatively impacted Conv10 RCA by −0.042 to −0.182. We also show that adding augmented data when the number of original samples is low has the greatest positive impact on accuracy and this effect was larger with ResNet50 models. Our work demonstrates that using data augmentation where few original samples are available can considerably improve model performance and highlights the potential of augmentation in developing acoustic classifiers for species where data are limited or difficult to obtain.

EASE: An enhanced active learning framework for aspect-based sentiment analysis based on sample diversity and data augmentation

Aspect-Based Sentiment Analysis (ABSA) has received considerable attention in recent studies. Powerful pre-trained models were proposed which can be fine-tuned for many Natural Language Processing (NLP) tasks, including ABSA. However, fine-tuning these models needs a relatively large amount of labeled data. In this research, we propose EASE; an active learning framework to minimize manual labeling effort. We extend the active learning technique by incorporating the concept of sample diversity where similar samples are not selected for labeling. Furthermore, we maximize the utility of these samples by incorporating data augmentation. EASE was evaluated on three benchmark ABSA datasets from three different domains. The results show that the reduction of the number of needed labeled samples ranges from 88% to 94% among the three datasets while maintaining accuracy. Our results show that active learning is an effective approach to reduce manual labeling effort while maintaining comparable performance. Moreover, it is possible to reduce the number of labeled data even further by incorporating sample diversity and data augmentation while maintaining performance.

Bootstrap each lead’s latent: A novel method for self-supervised learning of multilead electrocardiograms

Background and Objective:Electrocardiogram (ECG) is one of the most important diagnostic tools for cardiovascular diseases (CVDs). Recent studies show that deep learning models can be trained using labeled ECGs to achieve automatic detection of CVDs, assisting cardiologists in diagnosis. However, the deep learning models heavily rely on labels in training, while manual labeling is costly and time-consuming. This paper proposes a new self-supervised learning (SSL) method for multilead ECGs: bootstrap each lead’s latent (BELL) to reduce the reliance and boost model performance in various tasks, especially when training data are insufficient. Method:BELL is a variant of the well-known bootstrap your own latent (BYOL). The BELL aims to learn prior knowledge from unlabeled ECGs by pretraining, benefitting downstream tasks. It leverages the characteristics of multilead ECGs. First, BELL uses the multiple-branch skeleton, which is more effective in processing multilead ECGs. Moreover, it proposes intra-lead and inter-lead mean square error (MSE) to guide pretraining, and their fusion can result in better performances. Additionally, BELL inherits the main advantage of the BYOL: No negative pair is used in pretraining, making it more efficient. Results:In most cases, BELL surpasses previous works in the experiments. More importantly, the pretraining improves model performances by 0.69% ∼ 8.89% in downstream tasks when only 10% of training data are available. Furthermore, BELL shows excellent adaptability to uncurated ECG data from a real-world hospital. Only slight performance degradation occurs (<1% in most cases) when using these data. Conclusion:The results suggest that the BELL can alleviate the reliance on manual ECG labels from cardiologists, a critical bottleneck of the current deep learning-based models. In this way, the BELL can also help deep learning extend its application on automatic ECG analysis, reducing the cardiologists’ burden in real-world diagnosis.