High-quality Training Data Research Articles

Joint physics-based and data-driven time-lapse seismic inversion: Mitigating data scarcity

In carbon capture and sequestration, developing rapid and effective imaging techniques is crucial for real-time monitoring of the spatial and temporal dynamics of [Formula: see text] propagation during and after injection. With continuing improvements in computational power and data storage, data-driven techniques based on machine learning (ML) have been effectively applied to seismic inverse problems. In particular, ML helps alleviate the ill-posedness and high computational cost of full-waveform inversion (FWI). However, such data-driven inversion techniques require massive high-quality training data sets to ensure prediction accuracy, which hinders their application to time-lapse monitoring of [Formula: see text] sequestration. We develop an efficient “hybrid” time-lapse workflow that combines physics-based FWI and data-driven ML inversion. The scarcity of the available training data is addressed by developing a new data-generation technique with physics constraints. The method is validated using a synthetic [Formula: see text]-sequestration model based on the Kimberlina storage reservoir in California. Our approach is shown to synthesize a large volume of high-quality, physically realistic training data, which is critically important in accurately characterizing the [Formula: see text] movement in the reservoir. The developed hybrid methodology can also simultaneously predict the variations in velocity and saturation and achieve high spatial resolution in the presence of realistic noise in the data.

Generation of training data for named entity recognition of artworks

As machine learning techniques are being increasingly employed for text processing tasks, the need for training data has become a major bottleneck for their application. Manual generation of large scale training datasets tailored to each task is a time consuming and expensive process, which necessitates their automated generation. In this work, we turn our attention towards creation of training datasets for named entity recognition (NER) in the context of the cultural heritage domain. NER plays an important role in many natural language processing systems. Most NER systems are typically limited to a few common named entity types, such as person, location, and organization. However, for cultural heritage resources, such as digitized art archives, the recognition of fine-grained entity types such as titles of artworks is of high importance. Current state of the art tools are unable to adequately identify artwork titles due to unavailability of relevant training datasets. We analyse the particular difficulties presented by this domain and motivate the need for quality annotations to train machine learning models for identification of artwork titles. We present a framework with heuristic based approach to create high-quality training data by leveraging existing cultural heritage resources from knowledge bases such as Wikidata. Experimental evaluation shows significant improvement over the baseline for NER performance for artwork titles when models are trained on the dataset generated using our framework.

Applying XAI to an AI-based system for candidate management to mitigate bias and discrimination in hiring

Assuming that potential biases of Artificial Intelligence (AI)-based systems can be identified and controlled for (e.g., by providing high quality training data), employing such systems to augment human resource (HR)-decision makers in candidate selection provides an opportunity to make selection processes more objective. However, as the final hiring decision is likely to remain with humans, prevalent human biases could still cause discrimination. This work investigates the impact of an AI-based system’s candidate recommendations on humans’ hiring decisions and how this relation could be moderated by an Explainable AI (XAI) approach. We used a self-developed platform and conducted an online experiment with 194 participants. Our quantitative and qualitative findings suggest that the recommendations of an AI-based system can reduce discrimination against older and female candidates but appear to cause fewer selections of foreign-race candidates. Contrary to our expectations, the same XAI approach moderated these effects differently depending on the context.

Assisted Cement Log Interpretation Using Machine Learning

Summary The Assisted Cement Log Interpretation Project has used machine learning (ML) to create a tool that interprets cement logs by predicting a predefined set of annular condition codes used in the cement log interpretation process. The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals. By using high-quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time. The tool consists of a training and a prediction tool integrated with cased-hole logging interpretation software. By containerizing the code using an “API First” design principle (API: application programming interface), the applicability of this add-on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further. To improve cement log interpretation consistency in the industry, the code will be made available as open source.

Deep-learning-guided high-resolution subsurface reflectivity imaging with application to ground-penetrating radar data

SUMMARY Subsurface reflectivity imaging is one of the most important geophysical characterization methods for revealing subsurface structures. In many cases, accurate subsurface reflectivity imaging is challenging because of, for example, random or coherent noise in the data and sparse source–receiver observation geometry. We develop a deep-learning-guided iterative imaging method to improve subsurface structure imaging. Specifically, we train a supervised neural network to infer a noise-free, high-resolution image from a noisy, low-resolution image and use this estimated image as guidance to regularize least-squares imaging. We develop a systematic method to generate high-quality synthetic training data (data-label pairs) to train the guidance neural network. The trained neural network can provide high-fidelity predictions even for field-data images that are not in the training data. We validate our new imaging method using one synthetic and two field ground-penetrating radar data examples, and find that our method can produce clean, high-resolution subsurface reflectivity images where existing single-pass and least-squares imaging methods fail due to noise and insufficient data coverage.

Uncertainty-aware self-supervised neural network for liver T 1ρ mapping with relaxation constraint

Objective. mapping is a promising quantitative MRI technique for the non-invasive assessment of tissue properties. Learning-based approaches can map from a reduced number of weighted images but requires significant amounts of high-quality training data. Moreover, existing methods do not provide the confidence level of the estimation. We aim to develop a learning-based liver mapping approach that can map with a reduced number of images and provide uncertainty estimation. Approach. We proposed a self-supervised neural network that learns a mapping using the relaxation constraint in the learning process. Epistemic uncertainty and aleatoric uncertainty are modelled for the quantification network to provide a Bayesian confidence estimation of the mapping. The uncertainty estimation can also regularize the model to prevent it from learning imperfect data. Main results. We conducted experiments on data collected from 52 patients with non-alcoholic fatty liver disease. The results showed that when only collecting two -weighted images, our method outperformed the existing methods for quantification of the liver. Our uncertainty estimation can further regularize the model to improve the performance of the model and it is consistent with the confidence level of liver values. Significance. Our method demonstrates the potential for accelerating the mapping of the liver by using a reduced number of images. It simultaneously provides uncertainty of quantification which is desirable in clinical applications.

Efficient Deep Reinforcement Learning for Optimal Path Planning

In this paper, we propose a novel deep reinforcement learning (DRL) method for optimal path planning for mobile robots using dynamic programming (DP)-based data collection. The proposed method can overcome the slow learning process and improve training data quality inherently in DRL algorithms. The main idea of our approach is as follows. First, we mapped the dynamic programming method to typical optimal path planning problems for mobile robots, and created a new efficient DP-based method to find an exact, analytical, optimal solution for the path planning problem. Then, we used high-quality training data gathered using the DP method for DRL, which greatly improves training data quality and learning efficiency. Next, we established a two-stage reinforcement learning method where, prior to the DRL, we employed extreme learning machines (ELM) to initialize the parameters of actor and critic neural networks to a near-optimal solution in order to significantly improve the learning performance. Finally, we illustrated our method using some typical path planning tasks. The experimental results show that our DRL method can converge much easier and faster than other methods. The resulting action neural network is able to successfully guide robots from any start position in the environment to the goal position while following the optimal path and avoiding collision with obstacles.

Sample2Sample: an improved self-supervised denoising framework for random noise suppression in distributed acoustic sensing vertical seismic profile data

SUMMARYThe performance of supervised deep learning-based denoising methods relies on massive amounts of high-quality training data set with labels. However, data labelling is a time-consuming and tedious process, and the lack of labelled data set has become a major bottleneck affecting the development of supervised deep learning-based denoising methods. In recent years, denoising methods that only use unlabelled noisy data set for training have received more and more attention. Although these methods get rid of the dependence on labels, they usually have some specific requirements on the training data set. For example, the paired training data are required to be multiple noisy observations for each scene or obey a specific noise distribution, etc., which are often very challenging to meet in practical applications. In this study, we propose an improved self-supervised denoising framework based on Noise2Noise that only uses noisy seismic data set for training, and we name it sample2sample. The proposed denoising framework does not require multiple repeated acquisitions of seismic data to obtain multiple independent noisy observations for each scene used for training, and has no specific requirement for the noise distribution prior. Specifically, we introduce a random sampler to generate paired subsamples from some individual noisy seismic data for training. The corresponding elements in the two paired subsamples are adjacent in the original seismic data and approximately meet the training premise of Noise2Noise, that is the paired training data have the same signal. In addition, considering that there are some subtle differences in the signals of the paired subsamples generated by sampling, we also introduce a regularization loss to compensate for this. We conducted a qualitative and quantitative analysis of the denoising performance of the proposed method through further experiments, including synthetic data experiments and field data experiments.

SIMULATING LIDAR TO CREATE TRAINING DATA FOR MACHINE LEARNING ON 3D POINT CLOUDS

Abstract. 3D point clouds represent an essential category of geodata used in a variety of geoinformation applications. Typically, these applications require additional semantics to operate on subsets of the data like selected objects or surface categories. Machine learning approaches are increasingly used for classification. They operate directly on 3D point clouds and require large amounts of training data. An adequate amount of high-quality training data is often not available or has to be created manually. In this paper, we introduce a system for virtual laser scanning to create 3D point clouds with semantics information by utilizing 3D models. In particular, our system creates 3D point clouds with the same characteristics regarding density, occlusion, and scan pattern as those 3D point clouds captured in the real world. We evaluate our system with different data sets and show the potential to use the data to train neural networks for 3D point cloud classification.

Monitoring early-successional trees for tropical forest restoration using low-cost UAV-based species classification

Logged forests cover four million square kilometers of the tropics, capturing carbon more rapidly than temperate forests and harboring rich biodiversity. Restoring these forests is essential to help avoid the worst impacts of climate change. Yet monitoring tropical forest recovery is challenging. We track the abundance of early-successional species in a forest restoration concession in Indonesia. If the species are carefully chosen, they can be used as an indicator of restoration progress. We present SLIC-UAV, a new pipeline for processing Unoccupied Aerial Vehicle (UAV) imagery using simple linear iterative clustering (SLIC)to map early-successional species in tropical forests. The pipeline comprises: (a) a field verified approach for manually labeling species; (b) automatic segmentation of imagery into “superpixels” and (c) machine learning classification of species based on both spectral and textural features. Creating superpixels massively reduces the dataset's dimensionality and enables the use of textural features, which improve classification accuracy. In addition, this approach is flexible with regards to the spatial distribution of training data. This allowed us to be flexible in the field and collect high-quality training data with the help of local experts. The accuracy ranged from 74.3% for a four-species classification task to 91.7% when focusing only on the key early-succesional species. We then extended these models across 100 hectares of forest, mapping species dominance and forest condition across the entire restoration project.

Sparse Gaussian processes for multi-step motion prediction of space tumbling objects

Since the robotic capturing of a space tumbling object requires a rapidly accurate prediction of the object motion, this paper addresses the nonlinear multi-step motion prediction issue of the object based on machine learning theory and proposes a novel prediction approach, in which the prediction efficiency and the prediction accuracy are main focuses. To improve the prediction efficiency, sparse Gaussian processes are employed for the multi-step motion prediction, only providing the historical data of the object motion states. The sparse pseudo training data continuously optimized from real historical data makes it feasible for real-time predictions. To improve the prediction accuracy, a heuristic optimizer, Markov chain Monte Carlo, is integrated within sparse Gaussian processes to overcome the sensitivity problem caused by random initializations during the training process, thus avoids the local optimal training results. Then, a high-quality sparse pseudo training data is obtained to construct the predictive distributions during the testing process. The benchmark example and comparisons of several motion prediction applications illustrate the correctness and effectiveness of the proposed approach, and the experimental validation demonstrates its real-world potential.

Unsupervised knowledge-transfer for learned image reconstruction* *The work of RB is substantially supported by the i4health PhD studentship (UK EPSRC EP/S021930/1) and from The Alan Turing Institute (UK EPSRC EP/N510129/1), and that of ZK, SA and BJ by UK EPSRC EP/T000864/1, and that of SA and BJ also by UK EPSRC EP/V026259/1. AH acknowledges funding by Academy of Finland Projects

Deep learning-based image reconstruction approaches have demonstrated impressive empirical performance in many imaging modalities. These approaches usually require a large amount of high-quality paired training data, which is often not available in medical imaging. To circumvent this issue we develop a novel unsupervised knowledge-transfer paradigm for learned reconstruction within a Bayesian framework. The proposed approach learns a reconstruction network in two phases. The first phase trains a reconstruction network with a set of ordered pairs comprising of ground truth images of ellipses and the corresponding simulated measurement data. The second phase fine-tunes the pretrained network to more realistic measurement data without supervision. By construction, the framework is capable of delivering predictive uncertainty information over the reconstructed image. We present extensive experimental results on low-dose and sparse-view computed tomography showing that the approach is competitive with several state-of-the-art supervised and unsupervised reconstruction techniques. Moreover, for test data distributed differently from the training data, the proposed framework can significantly improve reconstruction quality not only visually, but also quantitatively in terms of PSNR and SSIM, when compared with learned methods trained on the synthetic dataset only.

Development of the LCMAP annual land cover product across Hawaiʻi

• Generated several training datasets by integrating existing datasets. • Produced annual land cover maps for the Hawaiian Islands at 30-m resolution over 20 years with different training datasets and Change detection and classification (CCDC) algorithm. • Summarized the effect of training datasets (e.g., spatial coverage, quality) on accuracies for different land cover types. • Revealed annual land cover changes in Hawaiʻi. Following the completion of land cover and change (LCC) products for the conterminous United States (CONUS), the U.S. Geological Survey's (USGS’s) Land Change Monitoring, Assessment, and Projection initiative has broadened the capability of characterizing continuous historical land change across the full Landsat records for Hawaiʻi at 30-meter resolution. One of the challenges of implementing the LCMAP framework to process annual land cover maps in Hawaiʻi is to collect sufficient high-quality training data. Although multiple datasets depicting land cover information are available in Hawaiʻi, they covered limited time frames and were produced from various remote sensing sources with different, classification categories, spatial resolution, and mapping accuracies. No solo product is suitable to provide LCMAP training data labels on its own. In this paper, we focused on enhancing the LCMAP training datasets to generate land cover products from 2000 to 2019 in Hawaiʻi. A total of 200 independent reference data plots were generated and manually interpreted for validating the mapping results produced by the training datasets. The results revealed that using the appropriate filter of multiple products as training data pools improved the classification model performance. The effect of training datasets (e.g., spatial coverage, quality) on accuracies for different land cover types were summarized. The LCMAP land surface change products for Hawaiʻi are available at https://doi.org/10.5066/P91E8M23 .

Adversarial Grammatical Error Generation: Application to Persian Language

Grammatical error correction (GEC) greatly benefits from large quantities of high-quality training data. However, the preparation of a large amount of labelled training data is time-consuming and prone to human errors. These issues have become major obstacles in training GEC systems. Recently, the performance of English GEC systems has drastically been enhanced by the application of deep neural networks that generate a large amount of synthetic data from limited samples. While GEC has extensively been studied in languages such as English and Chinese, no attempts have been made to generate synthetic data for improving Persian GEC systems. Given the substantial grammatical and semantic differences of the Persian language, in this paper, we propose a new deep learning framework to create large enough synthetic sentences that are grammatically incorrect for training Persian GEC systems. A modified version of sequence generative adversarial net with policy gradient is developed, in which the size of the model is scaled down and the hyperparameters are tuned. The generator is trained in an adversarial framework on a limited dataset of 8000 samples. Our proposed adversarial framework achieved bilingual evaluation understudy (BLEU) scores of 64.5% on BLEU-2, 44.2% on BLEU-3, and 21.4% on BLEU-4, and outperformed the conventional supervised-trained long short-term memory using maximum likelihood estimation and recently proposed sequence labeler using neural machine translation augmentation. This shows promise toward improving the performance of GEC systems by generating a large amount of training data.

Object classification on video data of meteors and meteor-like phenomena: algorithm and data

ABSTRACT Every moment, countless meteoroids enter our atmosphere unseen. The detection and measurement of meteors offer the unique opportunity to gain insights into the composition of our solar systems’ celestial bodies. Researchers therefore carry out a wide-area-sky-monitoring to secure 360-degree video material, saving every single entry of a meteor. Existing machine intelligence cannot accurately recognize events of meteors intersecting the earth’s atmosphere due to a lack of high-quality training data publicly available. This work presents four reusable open source solutions for researchers trained on data we collected due to the lack of available labelled high-quality training data. We refer to the proposed data set as the NightSkyUCP data set, consisting of a balanced set of 10 000 meteor- and 10 000 non-meteor-events. Our solutions apply various machine-learning techniques, namely classification, feature learning, anomaly detection, and extrapolation. For the classification task, a mean accuracy of 99.1 per cent is achieved. The code and data are made public at figshare with DOI 10.6084/m9.figshare.16451625.

A data-centric weak supervised learning for highway traffic incident detection

Using the data from loop detector sensors for near-real-time detection of traffic incidents on highways is crucial to averting major traffic congestion. While recent supervised machine learning methods offer solutions to incident detection by leveraging human-labeled incident data, the false alarm rate is often too high to be used in practice. Specifically, the inconsistency in the human labeling of the incidents significantly affects the performance of supervised learning models. To that end, we focus on a data-centric approach to improve the accuracy and reduce the false alarm rate of traffic incident detection on highways. We develop a weak supervised learning workflow to generate high-quality training labels for the incident data without the ground truth labels, and we use those generated labels in the supervised learning setup for final detection. This approach comprises three stages. First, we introduce a data preprocessing and curation pipeline that processes traffic sensor data to generate high-quality training data through leveraging labeling functions, which can be domain knowledge-related or simple heuristic rules. Second, we evaluate the training data generated by weak supervision using three supervised learning models—random forest, k-nearest neighbors, and a support vector machine ensemble—and long short-term memory classifiers. The results show that the accuracy of all of the models improves significantly after using the training data generated by weak supervision. Third, we develop an online real-time incident detection approach that leverages the model ensemble and the uncertainty quantification while detecting incidents. Overall, we show that our proposed weak supervised learning workflow achieves a high incident detection rate (0.90) and low false alarm rate (0.08).

Automatic Generation of Synthetic Colonoscopy Videos for Domain Randomization

Abstract An increasing number of colonoscopic guidance and assistance systems rely on machine learning algorithms which require a large amount of high-quality training data. In order to ensure high performance, the latter has to resemble a substantial portion of possible configurations. This particularly addresses varying anatomy, mucosa appearance and image sensor characteristics which are likely deteriorated by motion blur and inadequate illumination. The limited amount of readily available training data hampers to account for all of these possible configurations which results in reduced generalization capabilities of machine learning models. We propose an exemplary solution for synthesizing colonoscopy videos with substantial appearance and anatomical variations which enables to learn discriminative domain-randomized representations of the interior colon while mimicking real-world settings.

FedHGCDroid: An Adaptive Multi-Dimensional Federated Learning for Privacy-Preserving Android Malware Classification.

With the popularity of Android and its open source, the Android platform has become an attractive target for hackers, and the detection and classification of malware has become a research hotspot. Existing malware classification methods rely on complex manual operation or large-volume high-quality training data. However, malware data collected by security providers contains user privacy information, such as user identity and behavior habit information. The increasing concern for user privacy poses a challenge to the current malware classification scheme. Based on this problem, we propose a new android malware classification scheme based on Federated learning, named FedHGCDroid, which classifies malware on Android clients in a privacy-protected manner. Firstly, we use a convolutional neural network and graph neural network to design a novel multi-dimensional malware classification model HGCDroid, which can effectively extract malicious behavior features to classify the malware accurately. Secondly, we introduce an FL framework to enable distributed Android clients to collaboratively train a comprehensive Android malware classification model in a privacy-preserving way. Finally, to adapt to the non-IID distribution of malware on Android clients, we propose a contribution degree-based adaptive classifier training mechanism FedAdapt to improve the adaptability of the malware classifier based on Federated learning. Comprehensive experimental studies on the Androzoo dataset (under different non-IID data settings) show that the FedHGCDroid achieves more adaptability and higher accuracy than the other state-of-the-art methods.

Optical Remote Sensing Image Understanding With Weak Supervision: Concepts, methods, and perspectives

In recent years, supervised learning has been widely used in various tasks of optical remote sensing image understanding, including remote sensing image classification, pixel-wise segmentation, change detection, and object detection. The methods based on supervised learning need a large amount of high-quality training data and their performance highly depends on the quality of the labels. However, in practical remote sensing applications, it is often expensive and time-consuming to obtain large-scale data sets with high-quality labels, which leads to a lack of sufficient supervised information. In some cases, only coarse-grained labels can be obtained, resulting in the lack of exact supervision. In addition, the supervised information obtained manually may be wrong, resulting in a lack of accurate supervision. Therefore, remote sensing image understanding often faces the problems of incomplete, inexact, and inaccurate supervised information, which will affect the breadth and depth of remote sensing applications. In order to solve the above-mentioned problems, researchers have explored various tasks in remote sensing image understanding under weak supervision. This paper summarizes the research progress of weakly supervised learning in the field of remote sensing, including three typical weakly supervised paradigms: 1) Incomplete supervision, where only a subset of training data is labeled; 2) Inexact supervision, where only coarse-grained labels of training data are given; 3) Inaccurate supervision, where the labels given are not always true on the ground.

Low-quality training data detection method of EEG signals for motor imagery BCI system

BackgroundThe design and implementation of high-performance motor imagery-based brain computer interface (MI-BCI) requires high-quality training samples. However, fluctuation in subjects’ physiological and mental states as well as artifacts can produce the low-quality motor imagery electroencephalogram (EEG) signal, which will damage the performance of MI-BCI system. New methodIn order to select high-quality MI-EEG training data, this paper proposes a low-quality training data detection method combining independent component analysis (ICA) and weak classifier cluster. we also design and implement a new online BCI system based on motor imagery to verify the online processing performance of the proposed method. ResultIn order to verify the effectiveness of the proposed method, we conducted offline experiments on the public dataset called BCI Competition IV Data Set 2b. Furthermore, in order to verify the processing performance of the online system, we designed 60 groups of online experiments on 12 subjects. The online experimental results show that the twelve subjects can complete the system task efficiently (the best experiment is 135.6 s with 9 trials of subject S1). ConclusionThis paper demonstrated that the proposed low-quality training data detection method can effectively screen out low-quality training samples, so as to improve the performance of the MI-BCI system.