Source Datasets Research Articles

An Adaptive Transfer Learning Framework for Functional Classification

In this paper, we study the transfer learning problem in functional classification, aiming to improve the classification accuracy of the target data by leveraging information from related source datasets. To facilitate transfer learning, we propose a novel transferability function tailored for classification problems, enabling a more accurate evaluation of the similarity between source and target dataset distributions. Interestingly, we find that a source dataset can offer more substantial benefits under certain conditions than another dataset with an identical distribution to the target dataset. This observation renders the commonly-used debiasing step in the parameter-based transfer learning algorithm unnecessary under some circumstances to the classification problem. In particular, we propose two adaptive transfer learning algorithms based on the functional Distance Weighted Discrimination (DWD) classifier for scenarios with and without prior knowledge regarding informative sources. Furthermore, we establish the upper bound on the excess risk of the proposed classifiers, providing the statistical gain via transfer learning mathematically provable. Simulation studies are conducted to thoroughly examine the finite-sample performance of the proposed algorithms. Finally, we implement the proposed method to Beijing air-quality data, and significantly improve the prediction of the PM 2.5 level of a target station by effectively incorporating information from source datasets.

Cycle-Consistent Adversarial chest X-rays Domain Adaptation for pneumonia diagnosis

To deal with the challenge of limited annotated data for training target domain pneumonia classifier, domain adaptation methods leverage publicly available source dataset to enhance performance on the target domain. However, the existing methods lack effective feature representation to ensure a significant similarity between the source domain and the target domain. To address this problem, we propose a novel method, called Cycle-Consistent Adversarial chest X-rays Domain Adaptation (C2ADA) to diagnose pneumonia from Chest X-rays automatically, which can transfer knowledge from a publicly available large-scale source dataset to the small-scale target dataset and achieve high performance with fewer target domain samples. Specifically, C2ADA introduces the cycle-consistent adversarial module to enforce the feature extractor to learn domain-invariant features. Furthermore, unlike most existing domain adaptation methods that deal with the same tasks in the source domain and target domain, our perspective is that heterogeneous tasks can assist the target domain in learning more effectively. Therefore, C2ADA includes two subnetworks, one is for the source domain, which contains a multilabel classification task, and the other is for the target domain, which focuses on the binary classification task. To assess the effectiveness of our method, we conduct a comparative analysis with other domain adaptation approaches. The experimental results show that the proposed method can achieve 96.86% accuracy on the ChestXRay2017 dataset utilizing just 50 images.

Domain adaptation hyperspectral image fusion based on spatial-spectral domain separation

ABSTRACT Hyperspectral datasets captured by airborne are one of the important data sources in the field of hyperspectral image fusion. However, the limited number of data samples often hinders the optimal performance of deep learning-based methods on these datasets. Domain adaptation is expected to alleviate the issue of data scarcity. However, there are two issues that need to be addressed in applying domain adaptation to hyperspectral image fusion, which are the heterogeneity and spatial degradation diversity of hyperspectral data. In this paper, we propose a domain adaptation hyperspectral image fusion network based on spatial-spectral domain separation. The model, constructed with three-dimensional convolutional layers, effectively addresses the challenge of hyperspectral data heterogeneity. Compared with 2D convolution, 3D convolution is able to consider both spatial and spectral dimensions and extract spatial-spectral features efficiently. Based on this, we design a domain-separation architecture to extract domain-invariant and domain-private features in the spatial-spectral domain. The architecture achieves domain adaptation by extracting domain-invariant features from both datasets to obtain the prior knowledge in the source dataset and apply it to the target dataset. Additionally, the spatial degradation of different datasets varies due to distinct shooting conditions during the acquisition. Therefore, the fusion algorithms should be robust to various degradation scenarios. To tackle this challenge, we design the Observation Module to predict degradation information and the Degradation Information Modulation Module to apply it to the input, thereby enhancing the network’s robustness. Experiments on various datasets demonstrate that our method is qualitatively and quantitatively superior to state-of-the-art methods.

Temporal Protein Complex Identification Based on Dynamic Heterogeneous Protein Information Network Representation Learning.

Protein complexes, as the fundamental units of cellular function and regulation, play a crucial role in understanding the normal physiological functions of cells. Existing methods for protein complex identification attempt to introduce other biological information on top of the protein-protein interaction (PPI) network to assist in evaluating the degree of association between proteins. However, these methods usually treat protein interaction networks as flat homogeneous static networks. They cannot distinguish the roles and importance of different types of biological information, nor can they reflect the dynamic changes of protein complexes. In recent years, heterogeneous network representation learning has achieved great success in processing complex heterogeneous information and mining deep semantics. We thus propose a temporal protein complex identification method based on Dynamic Heterogeneous Protein information network Representation Learning, DHPRL. DHPRL naturally integrates multiple types of heterogeneous biological information in the cellular temporal dimension. It simultaneously models the temporal dynamic properties of proteins and the heterogeneity of biological information to improve the understanding of protein interactions and the accuracy of complex prediction. Firstly, we construct Dynamic Heterogeneous Protein Information Network (DHPIN) by integrating temporal gene expression information and GO attribute information. Then we design a dual-view collaborative contrast mechanism. Specifically, proposing to learn protein representations from two views of DHPIN (1-hop relation view and meta-path view) to model the consistency and specificity between nearest-neighbour bio information and deeper biological semantics. The dynamic PPI network is thereafter re-weighted based on the learned protein representations. Finally, we perform protein identification on the re-weighted dynamic PPI network. Extensive experimental results demonstrate that DHPRL can effectively model complicated biological information and achieve state-of-the-art performance in most cases.

Attention Retinex Network(A4R-Net) for face detection under low-light environment

The degradation of recognition rates in low-light environments is a critical issue in terms of security when using object and face recognition technologies in various locations. Existing low-light enhancement models have shown limitations in terms of computational cost and performance. However, this paper overcomes these limitations. The experimental results demonstrate that our model achieves the same performance as existing models with 13 times lower computational cost and a face detection performance of 82.2%.The structure of this paper is as follows: Introduction, which provides the background and explains the limitations of existing models. Proposed Method, which details the structure and working principles of A4R-Net. Experimental Results, which present the evaluation of low-light enhancement performance and the comparison of face detection using YOLOv4 [1]. Conclusion, which discusses the contributions of this research and future research directions.The source code and dataset is https://github.com/Obiru2698/obiru2698.github.io/

DTL-5G: Deep transfer learning-based DDoS attack detection in 5G and beyond networks

Network slicing is considered as a key enabler for 5G and beyond mobile networks for supporting a variety of new services, including enhanced mobile broadband, ultra-reliable and low-latency communication, and massive connectivity, on the same physical infrastructure. However, this technology increases the susceptibility of networks to cyber threats, particularly Distributed Denial-of-Service (DDoS) attacks. These attacks have the potential to cause service quality degradation by overloading network function(s) that are central to network slices to operate seamlessly. This calls for an Intrusion Detection System (IDS) as a shield against a wide array of DDoS attacks. In this regard, one promising solution would be the use of Deep Learning (DL) models for detecting possible DDoS attacks, an approach that has already made its way into the field given its manifest effectiveness. However, one particular challenge with DL models is that they require large volumes of labeled data for efficient training, which are not readily available in operational networks. A possible workaround is to resort to Transfer Learning (TL) approaches that can utilize the knowledge learned from prior training to a target domain with limited labeled data. This paper investigates how Deep Transfer Learning (DTL) based approaches can improve the detection of DDoS attacks in 5G networks by leveraging DL models, such as Bidirectional Long Short-Term Memory (BiLSTM), Convolutional Neural Network (CNN), Residual Network (ResNet), and Inception as base models. A comprehensive dataset generated in our 5G network slicing testbed serves as the source dataset for DTL, which includes both benign and different types of DDoS attack traffic. After learning features, patterns, and representations from the source dataset using initial training, we fine-tune base models using a variety of TL processes on a target DDoS attack dataset. The 5G-NIDD dataset, which has a sparse amount of annotated traffic pertaining to several DDoS attack generated in a real 5G network, is chosen as the target dataset. The results show that the proposed DTL models have performance improvements in detecting different types of DDoS attacks in 5G-NIDD dataset compared to the case when no TL is applied. According to the results, the BiLSTM and Inception models being identified as the top-performing models. BiLSTM indicates an improvement of 13.90%, 21.48%, and 12.22% in terms of accuracy, recall, and F1-score, respectively, whereas, Inception demonstrates an enhancement of 10.09% in terms of precision, compared to the models that do not adopt TL.

Can We Standardize Name Reconciliaton via OpenRefine?

Scientific names in biodiversity represent one of the oldest identifiers used in science. As a result, a common repetitive task is being able to reconcile a list of scientific names against curated data sources. Reconciliation allows one to determine if names in a list are spelled correctly, whether they are currently accepted, and their nomenclatural status. There are several online and local resources that provide reconciliation services. We share here the potential in interoperability across reconciliation tools. Global Names Verifier (GNverifier), Catalogue of Life, Global Biodiversity Information Facility (GBIF), Taxonomic Name Resolution Service (TNRS), LifeWatch, National Center for Biotechnology Information (NCBI), World Flora Online, Global Biotic Interactions (GloBI), Nomer, Wikidata, and others provide their own tools for name reconciliation. All these tools have their scope, design decisions, input, and output formats. It is often useful to do reconciliation using several such services, because they often include complementary data. However, with all the idiosyncrasies of services and lack of standardization, it is not an easy task (Islam et al. 2024). It would be great for researchers if all existing and future tools could be standardized. Then moving from one resource to another would be as easy as changing the URL. Implementing elements of Findable, Accessible, Interoperable, and Reusable (FAIR) data management principles would help to create such standards. However, standardizing all existing and future resources to a common interface would be difficult. Some of them have no monetary or programmatic means to modify their code, while others have more urgent priorities. Some resources support a specific research path where adhering to a rigid standard might hinder their innovation. In this paper we suggest interoperability between reconciliation tools by implementing the OpenRefine Reconciliation Service. OpenRefine is a popular and powerful reconciliation and data cleaning application. It is used by many researchers for data transformation and normalization. Any service that implements the OpenRefine Service can be incorporated into data-management workflows just by providing the service's OpenRefine-compatible URL. Such compatible services can easily be discovered by providing their metadata in the OpenRefine Services Registry. In this paper we discuss our implementation of the OpenRefine Service with the Global Names Verifier (GNverifier) reconciliation tool. GNverifier is developed at the Species File Group as a part of the Global Names Architecture initiative. It offers a powerful, configurable, fast way to reconcile scientific names. GNverifier software aggregates data from more than 100 source datasets. Queries return currently accepted names when provided in a dataset. It allows finding matches for names that historically had several suffixes and can do fuzzy and partial matches. It sorts data by many factors to reliably provide the best available results. With a strong focus on software optimization and a sophisticated matching algorithm, it can process 2000 names a second, making it one of the fastest services available. OpenRefine can use GNverifier directly because it is compatible with the OpenRefine protocol. As shown in Fig. 1, switching between GNverifier and Wikidata reconciliation of scientific names requires only change of a service URL. Implementation of the OpenRefine protocol might solve many standardization problems. Some resources already have it implemented (e.g., Wikidata, GNverifier Whitmire and Mozzherin 2023, WFO Plant List, IPNI). Many people already use OpenRefine for their other reconciliation needs. For them, the incorporation of name reconciliation would be especially beneficial because it will fit into their existing data-management workflow Fig. 2. Basic reconciliation by itself is standard by design and a big step forward. Beyond the basic reconciliation (as seen in Fig. 2) there are more data that researchers are interested in. The Service Protocol allows one to add optional "extended" features. For example, for scientific names, we provide "currently accepted" names, data sources where a name was found, taxonomic classification, etc. Fig. 3. To make these fields standardized, we would need a recommendation document that describes additional fields and their format. We need interested parties to participate in its creation and agree on its usage. The same would apply to optional input filters, for example for restricting reconciliation to certain data sources or higher taxonomic entities. We think OpenRefine would be a significant step forward for standardization between name-reconciliation tools.

Advancing mRNA subcellular localization prediction with graph neural network and RNA structure.

The asymmetrical distribution of expressed mRNAs tightly controls the precise synthesis of proteins within human cells. This non-uniform distribution, a cornerstone of developmental biology, plays a pivotal role in numerous cellular processes. To advance our comprehension of gene regulatory networks, it is essential to develop computational tools for accurately identifying the subcellular localizations of mRNAs. However, considering multi-localization phenomena remains limited in existing approaches, with none considering the influence of RNA's secondary structure. In this study, we propose Allocator, a multi-view parallel deep learning framework that seamlessly integrates the RNA sequence-level and structure-level information, enhancing the prediction of mRNA multi-localization. The Allocator models equip four efficient feature extractors, each designed to handle different inputs. Two are tailored for sequence-based inputs, incorporating multilayer perceptron and multi-head self-attention mechanisms. The other two are specialized in processing structure-based inputs, employing graph neural networks. Benchmarking results underscore Allocator's superiority over state-of-the-art methods, showcasing its strength in revealing intricate localization associations. The webserver of Allocator is available at http://Allocator.unimelb-biotools.cloud.edu.au; the source code and datasets are available on GitHub (https://github.com/lifuyi774/Allocator) and Zenodo (https://doi.org/10.5281/zenodo.13235798).

MetalTrans: A Biological Language Model-Based Approach for Predicting Disease-Associated Mutations in Protein Metal-Binding Sites.

The critical importance of accurately predicting mutations in protein metal-binding sites for advancing drug discovery and enhancing disease diagnostic processes cannot be overstated. In response to this imperative, MetalTrans emerges as an accurate predictor for disease-associated mutations in protein metal-binding sites. The core innovation of MetalTrans lies in its seamless integration of multifeature splicing with the Transformer framework, a strategy that ensures exhaustive feature extraction. Central to MetalTrans's effectiveness is its deep feature combination strategy, which merges evolutionary-scale modeling amino acid embeddings with ProtTrans embeddings, thus shedding light on the biochemical properties of proteins. Employing the Transformer component, MetalTrans leverages the self-attention mechanism to delve into higher-level representations. Utilizing mutation site information for feature fusion not only enriches the feature set but also sidesteps the common pitfall of overestimation linked to protein sequence-based predictions. This nuanced approach to feature fusion is a key differentiator, enabling MetalTrans to outperform existing methods significantly, as evidenced by comparative analyses. Our evaluations across varied metal binding site data sets (specifically Zn, Ca, Mg, and Mix) underscore MetalTrans's superior performance, which achieved the average AUC values of 0.971, 0.965, 0.980, and 0.945 on multiple 5-fold cross-validation, respectively. Remarkably, against the multichannel convolutional neural network method on a benchmark independent test set, MetalTrans demonstrated unparalleled robustness and superiority, boasting the AUC score of 0.998 on multiple 5-fold cross-validation. Our comprehensive examination of the predicted outcomes further confirms the effectiveness of the model. The source codes, data sets, and prediction results for MetalTrans can be accessed for academic usage at https://github.com/EduardWang/MetalTrans.

Metrics for Dataset Demographic Bias: A Case Study on Facial Expression Recognition.

Demographic biases in source datasets have been shown as one of the causes of unfairness and discrimination in the predictions of Machine Learning models. One of the most prominent types of demographic bias are statistical imbalances in the representation of demographic groups in the datasets. In this article, we study the measurement of these biases by reviewing the existing metrics, including those that can be borrowed from other disciplines. We develop a taxonomy for the classification of these metrics, providing a practical guide for the selection of appropriate metrics. To illustrate the utility of our framework, and to further understand the practical characteristics of the metrics, we conduct a case study of 20 datasets used in Facial Emotion Recognition (FER), analyzing the biases present in them. Our experimental results show that many metrics are redundant and that a reduced subset of metrics may be sufficient to measure the amount of demographic bias. The article provides valuable insights for researchers in AI and related fields to mitigate dataset bias and improve the fairness and accuracy of AI models.

Prediction of Super-enhancers Based on Mean-shift Undersampling

Background: Super-enhancers are clusters of enhancers defined based on the binding occupancy of master transcription factors, chromatin regulators, or chromatin marks. It has been reported that super-enhancers are transcriptionally more active and cell-type-specific than regular enhancers. Therefore, it is necessary to identify super-enhancers from regular enhancers. A variety of computational methods have been proposed to identify super-enhancers as auxiliary tools. However, most methods use ChIP-seq data, and the lack of this part of the data will make the predictor unable to execute or fail to achieve satisfactory performance. Objective: The aim of this study is to propose a stacking computational model based on the fusion of multiple features to identify super-enhancers in both human and mouse species. Methods: This work adopted mean-shift to cluster majority class samples and selected four sets of balanced datasets for mouse and three sets of balanced datasets for human to train the stacking model. Five types of sequence information are used as input to the XGBoost classifier, and the average value of the probability outputs from each classifier is designed as the final classification result. Results: The results of 10-fold cross-validation and cross-cell-line validation prove that our method has superior performance compared to other existing methods. The source code and datasets are available at https://github.com/Cheng-Han-max/SE_voting. Conclusion: The analysis of feature importance indicates that Mismatch accounts for the highest proportion among the top 20 important features.

AI4LUC: deep learning and automated mask labelling to support land use and land cover mapping in the Cerrado biome

ABSTRACT Land use and land cover (LULC) mapping in the Brazilian Cerrado is challenging due to its vastness, seasonal vegetation changes, and spectral class confusion. Artificial intelligence, particularly deep learning (DL), offers promising solutions through semantic segmentation of Cerrado images, though it traditionally requires manual image annotation. This paper introduces the AI4LUC (Artificial Intelligence for Land Use and Land Cover Classification) method, leveraging DL and automated mask labeling to address LULC in the Brazilian Cerrado. AI4LUC includes pre-processing modules and the Smart Mask©Labelling (SML) module, which combines morphological operations with the CerraNetv3 deep convolutional neural network to automatically generate training masks for DL models like U-Net and DeepLabv3+. Additionally, we present the Cerrado Dataset (CerraData-v3), a high-resolution image dataset with 80,000 labeled samples. Our experiments show the SML module achieved an F1-score of 0.6647 and mIoU of 0.5068, demonstrating significant potential for automated semantic segmentation in this large biome. Furthermore, DeepLabv3+ outperformed U-Net by 69.86%. Source code and dataset: https://github.com/ai4luc/AI4LUC.

MIGP: Metapath Integrated Graph Prompt Neural Network

Graph neural networks (GNNs) leveraging metapaths have garnered extensive utilization. Nevertheless, the escalating parameters and data corpus within graph pre-training models incur mounting training costs. Consequently, GNN models encounter hurdles including diminished generalization capacity and compromised performance amidst small sample datasets. Drawing inspiration from the efficacy demonstrated by self-supervised learning methodologies in natural language processing, we embark on an exploration. We endeavor to imbue graph data with augmentable, learnable prompt vectors targeting node representation enhancement to foster superior adaptability to downstream tasks. This paper proposes a novel approach, the Metapath Integrated Graph Prompt Neural Network (MIGP), which leverages learnable prompt vectors to enhance node representations within a pretrained model framework. By leveraging learnable prompt vectors, MIGP aims to address the limitations posed by mall sample datasets and improve GNNs’ model generalization. In the pretraining stage, we split symmetric metapaths in heterogeneous graphs into short metapaths and explicitly propagate information along the metapaths to update node representations. In the prompt-tuning stage, the parameters of the pretrained model are fixed, a set of independent basis vectors is introduced, and an attention mechanism is employed to generate task-specific learnable prompt vectors for each node. Another notable contribution of our work is the introduction of three patent datasets, which is a pioneering application in related fields. We will make these three patent datasets publicly available to facilitate further research on large-scale patent data analysis. Through comprehensive experiments conducted on three patent datasets and three other public datasets, i.e., ACM, IMDB, and DBLP, we demonstrate the superior performance of the MIGP model in enhancing model applicability and performance across a variety of downstream datasets. The source code and datasets are available in the website.11https://github.com/hzw-ai/MIGP, with password: MIGPNN2024.

Image-level supervision and self-training for transformer-based cross-modality tumor segmentation

Deep neural networks are commonly used for automated medical image segmentation, but models will frequently struggle to generalize well across different imaging modalities. This issue is particularly problematic due to the limited availability of annotated data, both in the target as well as the source modality, making it difficult to deploy these models on a larger scale. To overcome these challenges, we propose a new semi-supervised training strategy called MoDATTS. Our approach is designed for accurate cross-modality 3D tumor segmentation on unpaired bi-modal datasets. An image-to-image translation strategy between modalities is used to produce synthetic but annotated images and labels in the desired modality and improve generalization to the unannotated target modality. We also use powerful vision transformer architectures for both image translation (TransUNet) and segmentation (Medformer) tasks and introduce an iterative self-training procedure in the later task to further close the domain gap between modalities, thus also training on unlabeled images in the target modality. MoDATTS additionally allows the possibility to exploit image-level labels with a semi-supervised objective that encourages the model to disentangle tumors from the background. This semi-supervised methodology helps in particular to maintain downstream segmentation performance when pixel-level label scarcity is also present in the source modality dataset, or when the source dataset contains healthy controls. The proposed model achieves superior performance compared to other methods from participating teams in the CrossMoDA 2022 vestibular schwannoma (VS) segmentation challenge, as evidenced by its reported top Dice score of 0.87±0.04 for the VS segmentation. MoDATTS also yields consistent improvements in Dice scores over baselines on a cross-modality adult brain gliomas segmentation task composed of four different contrasts from the BraTS 2020 challenge dataset, where 95% of a target supervised model performance is reached when no target modality annotations are available. We report that 99% and 100% of this maximum performance can be attained if 20% and 50% of the target data is additionally annotated, which further demonstrates that MoDATTS can be leveraged to reduce the annotation burden.

Accuracy and transportability of machine learning models for adolescent suicide prediction with longitudinal clinical records

Machine Learning models trained from real-world data have demonstrated promise in predicting suicide attempts in adolescents. However, their transportability, namely the performance of a model trained on one dataset and applied to different data, is largely unknown, hindering the clinical adoption of these models. Here we developed different machine learning-based suicide prediction models based on real-world data collected in different contexts (inpatient, outpatient, and all encounters) with varying purposes (administrative claims and electronic health records), and compared their cross-data performance. The three datasets used were the All-Payer Claims Database in Connecticut, the Hospital Inpatient Discharge Database in Connecticut, and the Electronic Health Records data provided by the Kansas Health Information Network. We included 285,320 patients among whom we identified 3389 (1.2%) suicide attempters and 66% of the suicide attempters were female. Different machine learning models were evaluated on source datasets where models were trained and then applied to target datasets. More complex models, particularly deep long short-term memory neural network models, did not outperform simpler regularized logistic regression models in terms of both local and transported performance. Transported models exhibited varying performance, showing drops or even improvements compared to their source performance. While they can achieve satisfactory transported performance, they are usually upper-bounded by the best performance of locally developed models, and they can identify additional new cases in target data. Our study uncovers complex transportability patterns and could facilitate the development of suicide prediction models with better performance and generalizability.

Adaptive centroid prototype-based domain adaptation for fault diagnosis of rotating machinery without source data

Domain adaptation can effectively achieve fault diagnosis tasks with unlabeled target data using similar labeled source datasets during the training stage. However, the labeled source datasets are usually not directly accessible due to data privacy concerns, which restrict the application of the domain adaptation-based fault diagnosis methods. In this study, an adaptive centroid prototype-based domain adaptation (ACPDA) method is proposed to conduct fault diagnosis tasks in the unlabeled target domain without accessing source data. In ACPDA, an entropy-based adaptive prototype memory matrix is constructed to filter reliable samples and define the initial pseudo-label in the target domain. The centroid prototype is designed using all target data to update the pseudo-label and avoid confidence bias. Furthermore, the information maximization loss function is employed to reduce the feature distribution discrepancies. Extensive experiments on real wind turbine gearbox datasets demonstrate the effectiveness and superiority of the proposed ACPDA method.

Industrial Battery State-of-Health Estimation with Incomplete Limited Data Towards Second-Life Applications

Battery State of Health (SOH) estimation is vital across applications ranging from portable electronics to electric vehicles, particularly in second-life applications where accurate prediction becomes complex due to varying degradation levels. This paper introduces a novel SOH estimation model to address the lack of labeled data, employing Domain Adversarial Neural Networks (DANN) combined with 1-dimensional Convolutional Neural Networks (CNN). The proposed method allows for effective transfer of knowledge between diverse battery conditions, enhancing adaptability and efficiency by utilizing both source and target datasets. Experimental results demonstrate that the proposed model achieves a Mean Absolute Error (MAE) of 1.68% and a Root Mean Squared Error (RMSE) of 2.50%, with minimal data. Specifically, the model requires only one cell of unlabeled data from the second-life target domain, utilizing only the dQ/dV curve for estimation. Proposed model sets a new standard in second-life battery health monitoring and management by effectively leveraging a minimal amount of data for training, this approach offers a robust solution for accurate SOH estimation, particularly in scenarios with limited access to labeled data. Conflict of Interest Statement The authors declare no conflicts of interest.

CSP Data: A Data Discovery Web Application of Commercial CSP Plants

CSP Data is a free web application for data discovery of all commercial Concentrating Solar Power (CSP) plants. CSP data is fed with CSP.guru data, an open source dataset of all Concentrating Solar Power (CSP) plants. This paper introduces and explains the tools and information provided by CSP Data to potential users so they can use it efficiently together with the ongoing work being performed to release new features. It also presents the CSP.guru dataset, how the original idea was developed, its key features and how this dataset is used by industry and academia.

Distilling consistent relations for multi-source domain adaptive person re-identification

Although unsupervised person re-identification has made great progress by employing domain adaptation, existing methods rarely make use of multiple source datasets. These works neglect a practical scenario where annotated data from multiple domains are available. To address this issue, we propose a novel method to exploit the diversity and consistency of multiple sources to improve domain adaptation for person re-identification. In the proposed method, multiple expert models pre-trained in different source domains are adapted to the target domain by self-training with expert-specific clustering-based pseudo labels. To improve the consistency of multiple experts, we transfer knowledge between experts through dual similarity distillation between experts. Finally, to further encourage experts to learn complementary features, we maintain the feature diversity of multiple experts by representation decorrelation, preventing them from being too similar. Extensive experiments conducted on three benchmarks, Market-1501, DukeMTMC-reID, and MSMT17, demonstrate the effectiveness of the proposed method.

Transfer-Learning Prediction Model for Low-Cycle Fatigue Life of Bimetallic Steel Bars

The prediction of the low-cycle fatigue life of bimetallic steel bars (BSBs) is essential to promote the engineering application of BSBs. However, research on the low-cycle fatigue properties of BSB is limited, and fatigue experiments are time-consuming. Moreover, considering that sufficient data are needed for model training, the lack of data hinders the leverage of typical data-driven machine learning, which is widely used in fatigue life prediction. To address this issue, a transfer learning framework was suggested to accurately predict the low-cycle fatigue life of BSBs with limited data. To achieve this goal, 54 data points obtained from low-cycle fatigue tests on BSBs and 264 data points of other metallic bars were collected. Source models based on artificial neural networks (ANNs) were first constructed using the collected source dataset. Then, the learned knowledge stored in the source models was transferred to the transfer models. After that, transfer models were further fine-tuned and then tested using the target dataset of BSBs. The ANN models, which were of the same structure as the transfer models but only trained with the target dataset without transferring deep features from the source models, were set as baseline models. Compared with baseline models, the constructed transfer models could be used to accurately predict the fatigue life of BSBs. Moreover, the influence of hidden layers of ANNs on accuracy was examined by comparing one-layer and two-layer transfer models. Furthermore, the influence of key parameters on fatigue life of metallic bars was evaluated by feature analysis.