Search

In elderly monitoring or indoor intrusion detection, the recognition of human activity is a key task. Owing to several strengths of Wi-Fi-based devices, including their non-contact and privacy protection, these devices have been widely applied in the area of smart homes. By the deep learning technique, numerous Wi-Fi-based activity recognition methods can realize satisfied recognitions, however, these methods may fail to recognize the activities of an unknown person without the learning process. In this study, using channel state information (CSI) data, a novel cross-person activity recognition (CPAR) method is proposed by a deep learning approach with generalization capability. Combining one of the state-of-the-art deep neural networks (DNNs) used in activity recognition, i.e., attention-based bi-directional long short-term memory (ABLSTM), the snapshot ensemble is the first to be adopted to train several base-classifiers for enhancing the generalization and practicability of recognition. Second, to discriminate the extracted features, metric learning is further introduced by using the center loss, obtaining snapshot ensemble-used ABLSTM with center loss (SE-ABLSTM-C). In the experiments of CPAR, the proposed SE-ABLSTM-C method markedly improved the recognition accuracies to an application level, for seven categories of activities. [ABSTRACT FROM AUTHOR]

Background: The identification of compound-protein interactions (CPIs) is crucial for drug discovery and understanding mechanisms of action. Accurate CPI prediction can elucidate drug-target-disease interactions, aiding in the discovery of candidate compounds and effective synergistic drugs, particularly from traditional Chinese medicine (TCM). Existing in silico methods face challenges in prediction accuracy and generalization due to compound and target diversity and the lack of largescale interaction datasets and negative datasets for model learning. Methods: To address these issues, we developed a computational model for CPI prediction by integrating the constructed large-scale bioactivity benchmark dataset with a deep learning (DL) algorithm. To verify the accuracy of our CPI model, we applied it to predict the targets of compounds in TCM. An herb pair of Astragalus membranaceus and Hedyotis diffusaas was used as a model, and the active compounds in this herb pair were collected from various public databases and the literature. The complete targets of these active compounds were predicted by the CPI model, resulting in an expanded target dataset. This dataset was next used for the prediction of synergistic antitumor compound combinations. The predicted multi-compound combinations were subsequently examined through in vitro cellular experiments. Results: Our CPI model demonstrated superior performance over other machine learning models, achieving an area under the Receiver Operating Characteristic curve (AUROC) of 0.98, an area under the precision-recall curve (AUPR) of 0.98, and an accuracy (ACC) of 93.31% on the test set. The model's generalization capability and applicability were further confirmed using external databases. Utilizing this model, we predicted the targets of compounds in the herb pair of Astragalus membranaceus and Hedyotis diffusaas, yielding an expanded target dataset. Then, we integrated this expanded target dataset to predict effective drug combinations using our drug synergy prediction model DeepMDS. Experimental assay on breast cancer cell line MDA-MB-231 proved the efficacy of the best predicted multi-compound combinations: Combination I (Epicatechin, Ursolic acid, Quercetin, Aesculetin and Astragaloside IV) exhibited a half-maximal inhibitory concentration (IC50) value of 19.41 μM, and a combination index (CI) value of 0.682; and Combination II (Epicatechin, Ursolic acid, Quercetin, Vanillic acid and Astragaloside IV) displayed a IC50 value of 23.83 μM and a CI value of 0.805. These results validated the ability of our model to make accurate predictions for novel CPI data outside the training dataset and evaluated the reliability of the predictions, showing good applicability potential in drug discovery and in the elucidation of the bioactive compounds in TCM. Conclusion: Our CPI prediction model can serve as a useful tool for accurately identifying potential CPI for a wide range of proteins, and is expected to facilitate drug research, repurposing and support the understanding of TCM. [ABSTRACT FROM AUTHOR]

Copyright of Frontiers of Information Technology & Electronic Engineering is the property of Springer Nature and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The occurrence of class-imbalanced datasets is a frequent observation in natural science research, emphasizing the paramount importance of effectively harnessing them to construct highly accurate models for rockburst prediction. Initially, genuine rockburst incidents within a burial depth of 500 m were sourced from literature, revealing a small dataset imbalance issue. Utilizing various mainstream oversampling techniques, the dataset was expanded to generate six new datasets, subsequently subjected to 12 classifiers across 84 classification processes. The model incorporating the highest-scoring model from the original dataset and the top two models from the expanded dataset, yielded a high-performance model. Findings indicate that the KMeansSMOTE oversampling technique exhibits the most substantial enhancement across the combined 12 classifiers, whereas individual classifiers favor ET+SVMSMOTE and RF+SMOTENC. Following multiple rounds of hyper parameter adjustment via random cross-validation, the ET+SVMSMOTE combination attained the highest accuracy rate of 93.75%, surpassing mainstream models for rockburst prediction. Moreover, the SVMSMOTE technique, augmenting samples with fewer categories, demonstrated notable benefits in mitigating overfitting, enhancing generalization, and improving Recall and F1 score within RF classifiers. Validated for its high generalization performance, accuracy, and reliability. This process also provides an efficient framework for model development.

In this paper, the radial basis function (RBF) neural network was modified by data assimilation method to improve the modeling accuracy of high-dimensional aerodynamics. A correction factor γ was introduced into the kernel function of the traditional RBF neural network. The correction factor was corrected by using the EnKF data assimilation algorithm, and it was applied to the high-dimensional aerodynamic modeling prediction of the CRA309 rotor airfoil. The data assimilation method adopts a non-intrusive method to correct only the correction factor of the kernel function without destroying the overall architecture of the neural network, which greatly reduces the optimization parameters and variables, and significantly improves the modeling accuracy and efficiency of the RBF neural network. The modified RBF neural network model was then applied to high-dimensional aerodynamic modeling, and the aerodynamic parameters were predicted by simulation data instead. The design results verify the feasibility of the prediction model. In the case of less wind tunnel test data, it has certain engineering practical value to improve the utilization efficiency of test data.

Precision Agriculture (PA) has revolutionized crop management by leveraging information technology, satellite positioning data, and remote sensing. One crucial component in PA applications is the Normalized Difference Vegetation Index (NDVI), which offers valuable insights into crop vigor and health. However, discontinuity of optical satellite acquisitions related to cloud cover and the huge load of the required processing time pose challenges to real-time applications. NDVI prediction emerges as an innovative solution to address these limitations. It allows for proactive decision-making by providing accurate estimates, enabling farmers and land managers to plan essential agronomic activities such as irrigation, fertilization, and pest control, based on anticipated future conditions. This study introduces an Artificial Neural Network (ANN) model incorporating NDVI, Normalized Difference Water Index (NDWI), temperatures, and precipitation as predictive variables. The model employs a novel time series slicing algorithm, Boosting Adaptive Time Series Slicer (BATS), to enhance the input training dataset's variability, presenting the model with a broader range of examples. A 2-Bidirectional Long Short-Term Memory (LSTM) forecasting model was developed to predict future NDVI values over short and medium-term horizons. The study area used to train, test and validate the ANN corresponds to a diverse landscape of cultivated corn fields located in Piemonte (NW-Italy). Results showed that NDVI future estimates were accurate; considering three time horizons for predictions (5, 10, and 15 days) RMSE values resulted to be 0.028, 0.038 and 0.050, respectively. Additionally, ablation tests proved that the most important variable for enhancing the model's accuracy is the NDWI, and the most useful timesteps are the four most recent ones. To preliminary investigate the capability of the ANN to operate over a wider and different area it was applied over the entire Europe, using the LUCAS dataset as reference map to locate corn fields. Results show RMSE of 0.062, 0.083 and 0.105 for the 5, 10 and 15 days forecasting horizons, respectively. The methodology proposed in this paper can be a possible alternative to more ordinary approaches for NDVI forecasting that nowadays appears to be a fundamental step for a proactive precision agriculture where crop management can be significantly improved. Future developments should explore the use of sequence-to-sequence ANNs to predict the development of multiple spectral indices over multiple crop types simultaneously. [ABSTRACT FROM AUTHOR]

LiDAR has been an important sensor in autonomous driving systems. Compared to the measurements provided by a radar or camera. LiDAR can provide more precise geometric inforination and be fused with other types of senors to tackle various perception tasks in autonomous driving. Among these perception tasks. semantic segmentation oii LiDAR point clouds has received more and more research interest and achieved compelling results. However. there are still two unsolved issues. The first one is about minority classes caused by data imbalance. which is an inevitable problem in large-scale outdoor scenes. The minority classes. which are small in a scene und result in very few LiDAR points. can be important objects to be recognized for self-driving cars. e.g., pedestrians, motorcycles, traffic signs. In order to solve this class imbalance problem. we use U-Net++ architecture und dice loss to enhance the IoU score for the minority classes. The second issue is generalization capability on different LiDAR resolutions. Existing methods mostly need to be retrained to deal with data collected by LiDARs with different resolutions. We adopt KPConv as convolution operator to tackle this issue. With U-Net++ und dice loss. we get 5.1% mIoU improvement on SemanticKITTI. especially 9.5% rnIoU improvement of minority classes compared with baseline. Moreover, we show the generalization capability of our model with KPConv by training on 64-beam dataset and testing on 32-beam and 128-beam dataset. We obtain 3.3% mIoU improvement on 128-beam dataset and 1.9% mIoU improvement on 32-beam dataset. [ABSTRACT FROM AUTHOR]

From the perspective of input features, information can be divided into independent information and correlation information. Current neural networks mainly concentrate on the capturing of correlation information through connection weight parameters supplemented by bias parameters. This paper introduces feature-wise scaling and shifting (FwSS) into neural networks for capturing independent information of features, and proposes a new neural network FwSSNet. In the network, a pair of scale and shift parameters is added before each input of each network layer, and bias is removed. The parameters are initialized as 1 and 0, respectively, and trained at separate learning rates, to guarantee the fully capturing of independence and correlation information. The learning rates of FwSS parameters depend on input data and the training speed ratios of adjacent FwSS and connection sublayers, meanwhile those of weight parameters remain unchanged as plain networks. Further, FwSS unifies the scaling and shifting operations in batch normalization (BN), and FwSSNet with BN is established through introducing a preprocessing layer. FwSS parameters except those in the last layer of the network can be simply trained at the same learning rate as weight parameters. Experiments show that FwSS is generally helpful in improving the generalization capability of both fully connected neural networks and deep convolutional neural networks, and FWSSNets achieve higher accuracies on UCI repository and CIFAR-10. [ABSTRACT FROM AUTHOR]

Current single-model methods for fine-grained image classification suffer from insufficient generalisation ability, while multi-model fusion methods suffer from weight curing. The study suggests and experimentally tests a dynamic weight multi-model fusion strategy for transfer learning-based fine-grained picture classification. The results of the experiment showed that the suggested fusion model enhanced recognition accuracy by 1.33%, 1.19%, and 0.83% compared to the single model on the medical dataset and 3.25%, 1.34%, and 7.28% on the agronomy dataset, respectively. Furthermore, when compared to the comparison method, the models under the proposed method of the study improved recognition accuracy by 0.18%, 0.61%, 0.43%, and 0.43% on the medical dataset, and the experimental time consumed was 3.25 minutes less than that of the sum-of-maximum-probabilities method; however, the fusion models of the proposed method of the study had higher recognition accuracy than that of the comparison met Overall, the proposed dynamic weight multi-model fusion method for fine-grained image classification using migration learning has better performance and generalisation ability, which can improve performance while reducing time cost, and has higher application value for the actual fine-grained image classification task.

AbstractRapid spatial evaluation of seismic disaster after earthquake occurrence is required in disaster emergency rescue management, because of its importance in decreasing casualties and property losses. Among many categories of seismic disaster, evaluation of earthquake-affected population is of great significance to clarify the severity of earthquake disaster. For simple classic regression model, it is difficult to describe the strong nonlinear relationship between multiple influencing factors and earthquake disasters. In present study, an optimized BP neural network model considering spatial characteristic of influencing factors is proposed to evaluate the population distribution affected by earthquake. The correlation between earthquake-affected population and influencing factors is analysed using data of 2013 Ms7.0 Lushan earthquake. Ten influencing factors including elevation, slope angle, population density, per capita GDP, distance to fault, distance to river, NDVI, PGA, PGV, and distance to the epicentre, were classified into environmental and seismic factors. Correlation analysis revealed that per capita GDP and PGA factor had a stronger correlation with the earthquake-affected population. The earthquake-affected population was evaluated using a BP neural network by optimizing training samples considering spatial characteristics of per capita GDP and PGA factors. Different numbers of sample points, instead of a random distribution of sample points, were generated in areas with different value intervals of the influencing factors. The optimized samples improved the convergence speed and generalization capability of neuron network compared to random samples. The trained network was applied to the 2017 Ms7.0 Jiuzhaigou earthquake to verify its prediction accuracy. The MAE of the estimated earthquake-affected populations of different counties under Jiuzhaigou earthquake were 1.276 people/km2 using network model from optimized samples, smaller than the results of network model from random samples and linear regression model. The results indicate that BP neural network, which considers correlation characteristics of factors, has capability to evaluate spatial earthquake disaster.

In recent years, the method of building models based on machine learning has achieved good results in the field of bearing fault diagnosis. However, due to the complexity and variability of the actual working environment, the collected rolling bearing vibration data not only comes from different loads, but also contains noise data. The existing models are unable to adapt to all operating environments and their fault diagnosis capabilities are significantly reduced especially when the collected data is noisy. In order to achieve higher fault diagnosis accuracy and robustness under different work conditions, a new fault diagnosis model 1LWCNNLSTM (One-layer wide convolutional and long-short term memory network) is proposed, which is a hybrid model based on convolutional neural network (CNN) and long-short term memory network (LSTM). Firstly, the model extracts features from the raw data using a wide convolutional kernel to attenuate the effect of noise, then fuses the features extracted from different convolutional kernels to generate a new sequence, and finally uses LSTM to learn the features in the new sequence. The impact of the model parameters is analyzed through extensive experiments and the proposed model has higher diagnostic accuracy under mixed load and noise when compared with existing models. Further analyses of model classification details through visualization techniques and confusion matrices demonstrate the high usability of the model. The experimental results show that the model proposed has better load generalization capability and noise immunity for the vibration data coming from the complex working environments. [ABSTRACT FROM AUTHOR]

In the past decades, Deep Learning (DL) frameworks have demonstrated excellent performance in modeling nonlinear interactions and are a promising technique to move beyond physics-based models. In this context, super-resolution techniques may present an accurate approach as subfilter-scale (SFS) closure model for Large Eddy Simulations (LES) in premixed combustion. However, DL models need to perform accurately in a variety of physical regimes and generalize well beyond their training conditions. In this work, a super-resolution Generative Adversarial Network (GAN) is proposed as closure model for the unresolved subfilter-stress and scalar-flux tensors of the filtered reactive Navier–Stokes equations solved in LES. The model trained on a premixed methane/air jet flame is evaluated a-priori on similar configurations at different Reynolds and Karlovitz numbers. The GAN generalizes well at both lower and higher Reynolds numbers and outperforms existing algebraic models when the ratio between the filter size and the Kolmogorov scale is preserved. Moreover, extrapolation at a higher Karlovitz number is investigated indicating that the ratio between the filter size and the thermal flame thickness may not need to be conserved in order to achieve high correlation in terms of SFS field. Generalization studies obtained on substantially different flame conditions indicate that successful predictive abilities are demonstrated if the generalization criterion is matched. Finally, the reconstruction of a scalar quantity, different from that used during the training, is evaluated, revealing that the model is able to reconstruct scalar fields with large gradients that have not been explicitly used in the training. The a-priori investigations carried out assess whether out-of-sample predictions are even feasible in the first place, providing insights into the quantities that need to be conserved for the model to perform well between different regimes, and represent a crucial step toward future embedding into LES numerical solvers. [ABSTRACT FROM AUTHOR]

Abstract Facial expression is one form of communication which being non‐verbal in nature precedes verbal communication in both origin and conception. Most of the existing methods for Automatic Facial Expression Recognition (AFER) are mainly focused on global feature extraction assuming that all facial regions contribute equal amount of discriminative information to predict the expression class. The detection and localization of facial regions that have significant contribution to expression recognition and extraction of highly discriminative feature distribution from those regions are not fully explored. The key contributions of the proposed work are developing novel feature distribution upon combining the discriminative power of shape and texture feature; determining the contribution of facial regions and identifying the prominent facial regions that hold abstract and highly discriminative information for expression recognition. The shape and texture features taken into consideration are Local Phase Quantization (LPQ), Local Binary Pattern (LBP), and Histogram of Oriented Gradients (HOG). Multiclass Support Vector Machine (MSVM) is used while one versus one classification. The proposed work is implemented on CK+, KDEF, and JAFFE benchmark facial expression datasets. The recognition rate of the proposed work is 94.2% on CK+ and 93.7% on KDEF, which is significantly more than the existing handcrafted feature‐based methods.

The hot deformation behaviors of a Ti46Al2Cr2Nb alloy were investigated at strain rates of 0.001–0.1 s−1 and temperatures of 910–1060 °C. Under given deformation conditions, the activation energy of the TiAl alloy could be estimated as 319 kJ/mol. The experimental results were predicted by different predictive models including three constitutive models and three data-driven models. The most accurate data-driven model and constitutive model were an artificial neural network (ANN) and an Arrhenius type strain-compensated Sellars (SCS) model, respectively. In addition, the generalization capability of ANN model and SCS model was examined under different deformation conditions. Under known deformation conditions, the ANN model could accurately predict the flow stress of TiAl alloys at interpolated and extrapolated strains with a coefficient of determination (R2) greater than 0.98, while the R2 value of the SCS model was smaller than 0.5 at extrapolated strains. However, both ANN and SCS models performed poorly under new deformation conditions. A hybrid model based on the SCS model and ANN predictions was shown to have a wider generalization capability. The present work provides a comprehensive study on how to choose a predictive model for the flow stress of TiAl alloys under different conditions. [ABSTRACT FROM AUTHOR]

Copyright of Systems Engineering & Electronics is the property of Journal of Systems Engineering & Electronics Editorial Department and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The influence of geological development factors such as reservoir heterogeneity needs to be comprehensively considered in the determination of oil well production control strategy. In the past, many optimization algorithms are introduced and coupled with numerical simulation for well control problems. However, these methods require a large number of simulations, and the experience of these simulations is not preserved by the algorithm. For each new reservoir, the optimization algorithm needs to start over again. To address the above problems, two reinforcement learning methods are introduced in this research. A personalized Deep Q-Network (DQN) algorithm and a personalized Soft Actor-Critic (SAC)algorithm are designed for optimal control determination of oil wells. The inputs of the algorithms are matrix of reservoir properties, including reservoir saturation, permeability, etc., which can be treated as images. The output is the oil well production strategy. A series of samples are cut from two different reservoirs to form a dataset. Each sample is a square area that takes an oil well at the center, with different permeability and saturation distribution, and different oil-water well patterns. Moreover, all samples are expanded by using image enhancement technology to further increase the number of samples and improve the coverage of the samples to the reservoir conditions. During the training process, two training strategies are investigated for each personalized algorithm. The second strategy uses 4 times more samples than the first strategy. At last, a new set of samples is designed to verify the model’s accuracy and generalization ability. Results show that both the trained DQN and SAC models can learn and store historical experience, and push appropriate control strategies based on reservoir characteristics of new oil wells. The agreement between the optimal control strategy obtained by both algorithms and the global optimal strategy obtained by the exhaustive method is more than 95%. The personalized SAC algorithm shows better performance compared to the personalized DQN algorithm. Compared to the traditional Particle Swarm Optimization (PSO), the personalized models were faster and better at capturing complex patterns and adapting to different geological conditions, making them effective for real-time decision-making and optimizing oil well production strategies. Since a large amount of historical experience has been learned and stored in the algorithm, the proposed method requires only 1 simulation for a new oil well control optimization problem, which showing the superiority in computational efficiency.

Facial expression is one form of communication which being non‐verbal in nature precedes verbal communication in both origin and conception. Most of the existing methods for Automatic Facial Expression Recognition (AFER) are mainly focused on global feature extraction assuming that all facial regions contribute equal amount of discriminative information to predict the expression class. The detection and localization of facial regions that have significant contribution to expression recognition and extraction of highly discriminative feature distribution from those regions are not fully explored. The key contributions of the proposed work are developing novel feature distribution upon combining the discriminative power of shape and texture feature; determining the contribution of facial regions and identifying the prominent facial regions that hold abstract and highly discriminative information for expression recognition. The shape and texture features taken into consideration are Local Phase Quantization (LPQ), Local Binary Pattern (LBP), and Histogram of Oriented Gradients (HOG). Multiclass Support Vector Machine (MSVM) is used while one versus one classification. The proposed work is implemented on CK+, KDEF, and JAFFE benchmark facial expression datasets. The recognition rate of the proposed work is 94.2% on CK+ and 93.7% on KDEF, which is significantly more than the existing handcrafted feature‐based methods. [ABSTRACT FROM AUTHOR]

To address the problem caused by mixed pixels in MODIS images for high-resolution crop mapping, this paper presents a novel spatial–temporal deep learning-based approach for sub-pixel mapping (SPM) of different crop types within mixed pixels from MODIS images. High-resolution cropland data layer (CDL) data were used as ground references. The contributions of this paper are summarized as follows. First, we designed a novel spatial–temporal depth-wise residual network (ST-DRes) model that can simultaneously address both spatial and temporal data in MODIS images in efficient and effective manners for improving SPM accuracy. Second, we systematically compared different ST-DRes architecture variations with fine-tuned parameters for identifying and utilizing the best neural network architecture and hyperparameters. We also compared the proposed method with several classical SPM methods and state-of-the-art (SOTA) deep learning approaches. Third, we evaluated feature importance by comparing model performances with inputs of different satellite-derived metrics and different combinations of reflectance bands in MODIS. Last, we conducted spatial and temporal transfer experiments to evaluate model generalization abilities across different regions and years. Our experiments show that the ST-DRes outperforms the other classical SPM methods and SOTA backbone-based methods, particularly in fragmented categories, with the mean intersection over union (mIoU) of 0.8639 and overall accuracy (OA) of 0.8894 in Sherman County. Experiments in the datasets of transfer areas and transfer years also demonstrate better spatial–temporal generalization capabilities of the proposed method. [ABSTRACT FROM AUTHOR]

Technology development and decentralized operations create changes in conventional electric systems, where load modeling has been a challenge in dynamic analysis. Consequently, accurate dynamic load models are required to ensure the quality of the studies in current systems. This paper presents an automatic strategy based on clustering, classification, and optimization algorithms, to obtain the load models in the case of several system operating conditions. The obtained load models are helpful for the planning, operation, and protection of electric power systems. The proposed approach validation is performed using the IEEE 14-bus test system, where high performance is obtained. The average obtained cross-validation error for the load models assigned to the 13 clusters of disturbances is $5.36\times 10^{-3}$ . The cross-validation error is used as a tolerance value to determine when an online assigned load model is suitable to represent the measured disturbance. The proposed tests show the strategy’s capabilities of defining the load model online, making this approach suitable for field applications.

While input or output-perturbation-based adversarial training techniques have been exploited to enhance the generalization capability of a variety of nonfuzzy and fuzzy classifiers by means of dynamic regularization, their performance may perhaps be very sensitive to some inappropriate adversarial samples. In order to avoid this weakness and simultaneously ensure enhanced generalization capability, this work attempts to explore a novel knowledge adversarial attack model for the zero-order Tagaki–Sugeno–Kang (TSK) fuzzy classifiers. The proposed model is motivated by exploiting the existence of special knowledge adversarial attacks from the perspective of the human-like thinking process when training an interpretable zero-order TSK fuzzy classifier. Without any direct use of adversarial samples, which is different from input or output perturbation-based adversarial attacks, the proposed model considers adversarial perturbations of interpretable zero-order fuzzy rules in a knowledge-oblivion and/or knowledge-bias or their ensemble to mimic the robust use of knowledge in the human thinking process. Through dynamic regularization, the proposed model is theoretically justified for its strong generalization capability. Accordingly, a novel knowledge adversarial training method called KAT is devised to achieve promising generalization performance, interpretability, and fast training for zero-order TSK fuzzy classifiers. The effectiveness of KAT is manifested by the experimental results on 15 benchmarking UCI and KEEL datasets. [ABSTRACT FROM AUTHOR]

In this work, we propose a no-reference video quality assessment method, aiming to achieve high-generalization capability in cross-content, -resolution and -frame rate quality prediction. In particular, we evaluate the quality of a video by learning effective feature representations in spatial-temporal domain. In the spatial domain, to tackle the resolution and content variations, we impose the Gaussian distribution constraints on the quality features. The unified distribution can significantly reduce the domain gap between different video samples, resulting in more generalized quality feature representation. Along the temporal dimension, inspired by the mechanism of visual perception, we propose a pyramid temporal aggregation module by involving the short-term and long-term memory to aggregate the frame-level quality. Experiments show that our method outperforms the state-of-the-art methods on cross-dataset settings, and achieves comparable performance on intra-dataset configurations, demonstrating the high-generalization capability of the proposed method. The codes are released at https://github.com/Baoliang93/GSTVQA [ABSTRACT FROM AUTHOR]

Spatiotemporal fusion is a topical framework for solving the mutual restricted problem between the spatial and temporal resolution of satellite images. We pioneer an approach to replace similarity measurement steps in spatiotemporal fusion algorithms with convolutional neural networks (CNNs), building a bridge between weight function-based models and the learning-based models. Specifically, we propose a nonlocal form that separates the relational computation part from the value representation part, and construct the CNN-based similarity weight learning block for learning normalized weights. The block can be inserted into spatial and temporal adaptive reflectance fusion model (STARFM) to replace the manually designed weight calculation rules common in weight function-based methods, or into the CNN model StfNet to better utilize neighboring high-resolution images. The trained model outputs a high-resolution prediction from each base date image pair. The final result is a combination of the two predictions. In this regard, we propose the standard deviation-based weights to combine two prediction results. Four experiments are performed on Landsat–Moderate-resolution Imaging Spectroradiometer (MODIS) image pairs to determine the following: 1) the performance of the model at the target training date; 2) the generalization of the model in the target training time period; and 3) the generalization of the model at different dates and different geographical locations, each considering the different cases of giving one and two pairs of known images. Experimental results demonstrate the superiority of the similarity weight learning block and standard deviation-based weights. Among them, STARFM with the similarity weight learning block exhibits strong generalization, which testifies to the practical value of our model. [ABSTRACT FROM AUTHOR]

The deep learning based approaches which have been repeatedly proven to bring benefits to visual recognition tasks usually make a strong assumption that the training and test data are drawn from similar feature spaces and distributions. However, such an assumption may not always hold in various practical application scenarios on visual recognition tasks. Inspired by the hierarchical organization of deep feature representation that progressively leads to more abstract features at higher layers of representations, we propose to tackle this problem with a novel feature learning framework, which is called GMFAD, with better generalization capability in a multilayer perceptron manner. We first learn feature representations at the shallow layer where shareable underlying factors among domains (e.g., a subset of which could be relevant for each particular domain) can be explored. In particular, we propose to align the domain divergence between domain pair(s) by considering both inter-dimension and inter-sample correlations, which have been largely ignored by many cross-domain visual recognition methods. Subsequently, to learn more abstract information which could further benefit transferability, we propose to conduct feature disentanglement at the deep feature layer. Extensive experiments based on different visual recognition tasks demonstrate that our proposed framework can learn better transferable feature representation compared with state-of-the-art baselines. [ABSTRACT FROM AUTHOR]

• SSL-KNN is proposed to effectively classify stylistic and ordinary data. • Shared stylistic information is first introduced in SSL-KNN on stylistic data. • Gaussian kernel function is utilized in SSL-KNN for ordinary data. • The effectiveness of shared stylistic information is experimentally justified. While previous linear k nearest neighbor methods have demonstrated strong generalization capability for ordinary datasets, they often fail to achieve comparable generalization capability when applied to stylistic datasets due to the unique characteristics of this type of datasets. Additionally, when multiple testing samples share a certain style characteristic, both the traditional k nearest neighbor method and evolved versions of the linear k nearest neighbor method tend to ignore the shared style in the testing sample set, leading to inefficient utilization of information resources. In order to address these issues, we propose a new classification method called S hared S tyle L inear k N earest N eighbor (SSL-KNN), which enhances the generalization capability of the linear k nearest neighbor method on stylistic datasets. The proposed method is motivated by the observation that one can better determine the class of a collection of homogeneous samples by evaluating their overall style, rather than simply aggregating individual sample judgments. The proposed method achieves this goal and two additional goals: a) we obtain the style membership vectors of the testing samples with different classes of the training samples in order to seek a guarantee of the generalization capability on the stylistic datasets; b) we introduce a Gaussian kernel distance metric constraining the linear expression weights in order to seek a guarantee of the generalization capability on the ordinary datasets. Furthermore, we propose an alternating optimization strategy implemented to optimize the proposed SSL-KNN method. Finally, we evaluate the proposed method and the comparative methods on 15 benchmark datasets containing both ordinary and stylistic datasets. The results show that the proposed SSL-KNN method guarantees good generalization capability on both ordinary and stylistic datasets. [ABSTRACT FROM AUTHOR]

Electric microgrids require accurate dynamic models for operation, control, stability, and protection studies, then adequate load modeling plays an important role. This paper presents a two-stage adaptive approach to improve the generalization capability of load models obtained with the measurement-based modeling. The load model and their respective parameters are obtained through machine learning tools like decision trees (DTs) and optimization algorithms as ant colony (ACO). In the off-line stage of the proposed approach, several parameterized load models are optimally obtained using a database of microgrid disturbances. Then, the best model to represent each disturbance is defined using a similarity criterion. This model and the disturbance characteristics are integrated into a DT (classifier), while the characteristics and the model parameters are related in a second DT (predictor). These DTs are used in an on-line stage to swiftly determine the adequate parameterized load model in the case of a new disturbance in the microgrid. The approach’s performance is compared with the conventional measurement-based load modeling in a modified CIGRE benchmark low voltage microgrid. The results evidence the advantages of the proposed adaptive approach for dynamic load modeling. [ABSTRACT FROM AUTHOR]

There has been an increasing consensus in learning based face anti-spoofing that the divergence in terms of camera models is causing a large domain gap in real application scenarios. We describe a framework that eliminates the influence of inherent variance from acquisition cameras at the feature level, leading to the generalized face spoofing detection model that could be highly adaptive to different acquisition devices. In particular, the framework is composed of two branches. The first branch aims to learn the camera invariant spoofing features via feature level decomposition in the high frequency domain. Motivated by the fact that the spoofing features exist not only in the high frequency domain, in the second branch the discrimination capability of extracted spoofing features is further boosted from the enhanced image based on the recomposition of the high-frequency and low-frequency information. Finally, the classification results of the two branches are fused together by a weighting strategy. Experiments show that the proposed method can achieve better performance in both intra-dataset and cross-dataset settings, demonstrating the high generalization capability in various application scenarios. [ABSTRACT FROM AUTHOR]

Pu-erh tea processed from the sun-dried green tea leaves can be divided into ancient tea (AT) and terrace tea (TT) according to the source of raw material. However, their similar appearance makes AT present low market identification, resulting in a disruption in the tea market rules of fair trade. Therefore, this study analyzed the classification by principal component analysis/hierarchical clustering analysis and conducted the discriminant model through stepwise Fisher discriminant analysis and decision tree analysis based on the contents of water extract, phenolic components, alkaloid, and amino acids, aiming to investigate whether phytochemicals coupled with chemometric analyses distinguish AT and TT. Results showed that there were good separations between AT and TT, which was caused by 16 components with significant (p < 0.05) differences. The discriminant model of AT and TT was established based on six discriminant variables including water extract, (+)-catechin, (−)-epicatechin, (−)-epigallocatechin, theacrine, and theanine. Among them, water extract comprised multiple soluble solids, representing the thickness of tea infusion. The model had good generalization capability with 100% of performance indexes according to scores of the training set and model set. In conclusion, phytochemicals coupled with chemometrics analyses are a good approach for the identification of different raw materials.

In this article, we propose a face spoofing detection method by learning to fuse high-frequency (HF) and low-frequency (LF) features, in an effort to improve the generalization capability and fill up the domain gap between training and testing when the antispoofing is practically conducted in unseen scenarios. In particular, the proposed face antispoofing model consists of two streams that extract HF and LF components of a facial image with three high-pass and three low-pass filters. Moreover, considering the fact that spoofing features exist in different feature levels, we train our network with a novel multiscale triplet loss. The cross-frequency spatial attention module further enables the two streams to communicate and exchange information with each other. Finally, the outputs of the two streams are fused with a weighting strategy for final classification. Extensive experiments conducted on intra- and cross-database settings show the superiority of the proposed scheme. [ABSTRACT FROM AUTHOR]

Stochastic configuration network (SCN) that randomly assigns the weights connecting the input layer and the hidden layer with an inequality constraint can achieve a fast learning speed for dealing with regression tasks. In this paper, a driving amount based SCN (DASCN) is proposed to improve the performance in terms of generalization and structure compactness, which have gained considerable attention in the industrial process. In the proposed DASCN, the driving amount incorporated into SCN is used to further improve the structural parameters, especially the output weights, of SCN. The performance of DASCN is evaluated by function approximation, four benchmark datasets and practical application in the industrial process. The simulation results indicate that the DASCN has better generalization capability and a more compact network structure compared to other methods. [ABSTRACT FROM AUTHOR]

Recently, rain streaks removal from a single image has attracted much research attention to alleviate the degenerated performance of computer vision tasks implemented on rainy images. In this paper, we provide a thorough review for current single-image-based rain removal techniques, which can be mainly categorized into three classes: early filter-based, conventional prior-based, and recent deep learning-based approaches. Furthermore, inspired by the rationality of current deep learning-based methods and insightful characteristics underlying rain shapes, we build a specific coarse-to-fine deraining network architecture, which can finely deliver the rain structures and progressively removes rain streaks from the input image, accordingly. The superiority of the proposed network is substantiated by experiments implemented on synthetic and real rainy images both visually and quantitatively, as compared with comprehensive state-of-the-art methods along this line. Especially, it is verified that the proposed network possesses better generalization capability on real rainy images, implying its potential usefulness for this task. [ABSTRACT FROM AUTHOR]

This paper proposes the matrix modular neural network (MMNN), which is a modular neural network and adopts a novel task decomposition technique to solve complex problems, such as the large training sets and the category asymmetric training sets. A complex problem can be decomposed into many easier problems, each of which is dealt in two subspaces and can be solved by a single neural network module. All of these modules form a neural network matrix, which produces an output matrix that leads to an integration machine so that finally a classification decision result can be efficiently made. This paper's theoretic analyses and experiments show that the MMNN can reduce the learning time and improve the generalization capability and the classification accuracy of neural networks. [ABSTRACT FROM AUTHOR]

This paper aims to present a method for improving the modeling precision of static friction. To some extent, as the traditional static friction models available can't be unified to characterize all the friction situations, a back propagation neural network (BPNN) was proposed to weaken the requirements of traditional static friction models. In details, relative speed of interacting surfaces and joint load are typically considered as the inputs of BPNN, whose output is the predicted static friction. Furthermore, to speed up the convergence and improve the global generalization capability of BPNN, we use genetic algorithm (GA) to optimize the initial values of weights and thresholds. All the training samples follow with reciprocating constant-speed experiments of friction under the changes of joint speed and load. Three comparative experiments indicate that using GA to optimize the initial values of weights and thresholds benefit to improve the convergence rate of network and prediction accuracy, and comparing with the traditional model of static friction, the BPNN model has a higher prediction precision and excellent generalization capability. [ABSTRACT FROM AUTHOR]

The widespread adoption of photovoltaic (PV) technology for renewable energy necessitates accurate segmentation of PV panels to estimate installation capacity. However, achieving highly efficient and precise segmentation methods remains a pressing challenge. Recent advancements in artificial intelligence and remote sensing techniques have shown promise in PV segmentation. Nevertheless, real-world scenarios introduce complexities such as diverse sensing platforms, sensors, panel categories, and testing regions. These factors contribute to resolution, size, and foreground-background class imbalances, impeding accurate and generalized PV panel segmentation over large areas. To address these challenges, we propose GenPV, a deep learning model that leverages data distribution analysis and PV panel characteristics to enhance segmentation accuracy and generalization. GenPV employs a multi-scale feature learning approach, utilizing an enhanced feature pyramid network to fuse data features from multiple resolutions, effectively addressing resolution imbalance. Moreover, inductive learning is employed through a multitask approach, facilitating the detection and identification of both small and large-sized PV panels to mitigate size imbalance. To address significant class imbalance in PV panel recognition tasks, we integrate the Focal loss function for effective hard sample mining. Through experimental evaluation conducted in Heilbronn, Germany, our proposed method demonstrates superior performance compared to state-of-the-art approaches in PV panel segmentation. The results exhibit progressively higher accuracy and improved generalization capability. These findings highlight the potential of our method to serve as an advanced and practical tool for PV segmentation in the renewable energy field. • Analysis of the imbalance problems in real-world PV semantic segmentation scenarios. • Creation of an open-source PV panel dataset to advance renewable energy development. • Introducing a novel end-to-end DL model named GenPV for PV panel segmentation. • Improved accuracy and generalization in PV segmentation across unaligned datasets. [ABSTRACT FROM AUTHOR]

The spatiotemporal development of turbulent pipe flows with various conditions of pulsation is predicted by deep learning. In the present model, convolutional autoencoder (CAE) extracts the spatial features of the input as latent space vectors and long short-term memory (LSTM) evolves them in time. Time-delay neural network (TDNN) is accompanied by LSTMs to treat general unsteady variations of the spatial mean pressure gradient of the pulsatile flows. Temporal evolution is processed in a sequence-to-sequence manner. Note that the pulsating turbulent pipe flows are statistically unsteady flows for which the spatiotemporal development has been seldom predicted by the deep learning technique. The flow field obtained by direct numerical simulation (DNS) was used to train the model. When the pulsating parameters for the training and prediction are the same, several model parameters are validated. To examine further generalization capability, the present model was trained by data from four pulsation conditions. The model was tested afterward by test data from twenty different pulsation conditions of which the range of parameters is either within or out of the range covered by the training data. The model successfully predicts the various pulsating flow fields with reasonable accuracy. The change in the period yields a larger influence on prediction accuracy than the change in the amplitude. The degradation of the prediction accuracy of the bulk velocity for the case with a long period and large amplitude is not caused by the latent space vectors themselves, but by the decoder which reconstructs the flow field from the latent space vectors. • Development of pulsating turbulent pipe flow is predicted by deep learning. • The present model consists of CAE and LSTM coupled with TDNN. • The model predicts various pulsating flows under different conditions from that for the training. [ABSTRACT FROM AUTHOR]

Extreme learning machines (ELMs) are commonly adopted as base learners in ensemble methods due to unstable results and fast learning speed. However, most existing ensemble structures require explicit aggregation of ELM base learners' results before the final decision. This study proposes a novel Biological Ensemble ELMs called BE-ELM from a biological perspective for the first time and exploit the superiority of assembling multiple ELM base learners in parallel without the need for explicit aggregation, and thus simplifying the learning procedure. BE-ELM's structure is inspired by recent MIT neuroscience findings of brain learning mechanisms. The expression of the analytical solution of BE-ELM is similar to that of basic ELMs, which allows it to inherit their fast learning speed in the ensemble structure. Moreover, BE-ELM also has the added advantage of superior performance. Our theoretical analysis shows that BE-ELM consisting of multiple ELM base learners built on subsets of the original features is equivalent to a more complex ELM on the original feature space. We prove that BE-ELM, even without any explicit aggregation, can guarantee enhanced generalization capabilities. To confirm our findings, we conducted extensive experiments on various datasets. The results demonstrate that BE-ELM outperforms traditional ELMs and other state-of-the-art ensemble ELMs in terms of generalization performance on most datasets. These findings suggest that BE-ELM has the potential to improve prediction outcomes in practical applications. • Unlike other ensemble methods, BE-ELM requires no explicit aggregation. • BE-ELM has a clear biological analogy. • BE-ELM has a guarantee of enhanced generalization capability. • BE-ELM outperforms the standard ELMs and other state-of-art ensemble methods. • The proposed biological structure can be extended to different learning algorithms. [ABSTRACT FROM AUTHOR]

Extreme learning machine (ELM) is a non-iterative algorithm for training single-hidden layer feed-forward networks, whose training speed is much faster than those of conventional neural networks. However, its objective is only to minimize the empirical risk, which may cause overfitting easily. To overcome this defect, this paper proposes a novel algorithm named LGE2LM based on the localized generalization error model which provides an upper bound of generalization error through adopting the stochastic sensitivity. LGE2LM aims to improve the generalization capability of ELM by using the regularization technique which makes a trade-off between the empirical risk and the stochastic sensitive measure. The essence of LGE2LM is a quadratic problem without iterative process. Similar to ELM, all the parameters of this new algorithm are obtained without tuning, which makes the efficiency of LGE2LM much higher than those of traditional neural networks. Several experiments conducted on both artificial and real-world datasets show that LGE2LM has much better generalization capability and stronger robustness in comparison with ELM. [ABSTRACT FROM AUTHOR]