Unified Feature Space Research Articles

Harnessing structural stochasticity in the computational discovery and design of microstructures

This paper presents a deep generative model-based design methodology for tailoring the structural stochasticity of microstructures. Although numerous methods have been established for designing deterministic (periodic) or stochastic microstructures, a systematic design approach that allows the unified treatment of both deterministic and stochastic microstructure design domains has yet to be created. The proposed methodology resolves this issue by learning a unified feature space that embodies diverse structural patterns with continuously varying stochasticity levels. A highly diverse microstructure database is established to incorporate various types of deterministic and stochastic microstructure patterns. A property-aware deep generative model is proposed to learn a unified feature space of the structural characteristics, as well as the relationship between structure features and properties of interest. Autoencoder (AE), Variational Autoencoder (VAE), and Adversarial Autoencoder (AAE) are compared to understand their relative merits in the property-aware learning of the unified feature space. Microstructural designs with tailorable stochasticity and properties are obtained by searching the unified feature space. Multiple design cases are presented to demonstrate the capability of designing microstructures for structural stochasticity and properties. Furthermore, the proposed method is employed to create stochastically graded structures, which manipulate the mechanical behaviors by varying the local stochasticity of the structure.

DLI-Net: Dual Local Interaction Network for Fine-Grained Sketch-Based Image Retrieval

Fine-grained sketch-based image retrieval (FG-SBIR) is considered an ideal method of image retrieval due to the rich and easily accessible characteristics of sketches. It aims to find the most similar photo from the photo gallery based on the input sketch. Most previous works follow the paradigm that extracting global feature first and then projecting the features of sketch and photo to unified embedding feature space using triplet loss. However, the global feature is not appropriate for extracting the crucial fine-grained information. Based on this principle, we propose a Dual Local Interaction Network (DLI-Net). DLI-Net explores an effective and efficient way to utilize local features for FG-SBIR. Specifically, we first propose a Local Feature Extractor to extract mid-level local features. Then, in response to the problems brought by local features, we propose a Dual Interaction Module, which contains Self Interaction Module and Cross Interaction Module. Self Interaction Module speeds up retrieval by eliminating the redundant local features of background. Cross Interaction Module solves the spatial misalignment by making the sketches interact with photos. Extensive experiments on six commonly used datasets show that our DLI-Net outperforms state-of-the-art competitors by a significant margin with a reasonable retrieval speed. Moreover, to the best of our knowledge, DLI-Net is the first model that beats humans on all six datasets. Besides, DLI-Net also performs best on cross-category fine-grained sketch-based image retrieval task, which further demonstrates local features are more appropriate for FG-SBIR.

Distributed contrastive learning for medical image segmentation.

Supervised deep learning needs a large amount of labeled data to achieve high performance. However, in medical imaging analysis, each site may only have a limited amount of data and labels, which makes learning ineffective. Federated learning (FL) can learn a shared model from decentralized data. But traditional FL requires fully-labeled data for training, which is very expensive to obtain. Self-supervised contrastive learning (CL) can learn from unlabeled data for pre-training, followed by fine-tuning with limited annotations. However, when adopting CL in FL, the limited data diversity on each site makes federated contrastive learning (FCL) ineffective. In this work, we propose two federated self-supervised learning frameworks for volumetric medical image segmentation with limited annotations. The first one features high accuracy and fits high-performance servers with high-speed connections. The second one features lower communication costs, suitable for mobile devices. In the first framework, features are exchanged during FCL to provide diverse contrastive data to each site for effective local CL while keeping raw data private. Global structural matching aligns local and remote features for a unified feature space among different sites. In the second framework, to reduce the communication cost for feature exchanging, we propose an optimized method FCLOpt that does not rely on negative samples. To reduce the communications of model download, we propose the predictive target network update (PTNU) that predicts the parameters of the target network. Based on PTNU, we propose the distance prediction (DP) to remove most of the uploads of the target network. Experiments on a cardiac MRI dataset show the proposed two frameworks substantially improve the segmentation and generalization performance compared with state-of-the-art techniques.

A Fully Automatic Feature-Based Real-Time Traffic Surveillance System Using Data Association in the Probabilistic Framework

Multi-object tracking involves maintaining several trajectories of different objects moving in the scene throughout the video. With this objective, in this article, a fully automatic and real-time tracking algorithm to track multiple vehicles in a video is proposed. The proposed method specifically tries to address the challenges of occlusion and fast-motion in traffic surveillance. The algorithm begins with automatic detection of the moving targets by an adaptive GMM-based background subtraction method. The trajectories of these detected targets are then built using a three-level multi-motion modeled particle filter framework which allows to deal with the challenges of occlusion and fast-motions of the target. The likelihood model for targets is based on their color distribution and edge oriented histogram features. It is contended that the color distribution feature, which can represent the target appearance and the edge oriented histogram, which can describe the target structure are sufficient to represent it in an unique feature space. Based on the similarity of target likelihood, the locations of the targets are filtered. These filtered locations are then associated with the most likely detections using the proposed low-cost and fast data association algorithm based on Euclidean distance and prevailing motion vector. The performance evaluation of the proposed scheme is carried out based on the six measures: Multi Object Tracking Accuracy, Multi Object Tracking Precision, Mostly Tracked trajectories, Mostly Lost trajectories, Identity Switches and Frames Per Second. The results evaluated on stationary camera shot sequences from benchmark datasets as well as real-time shot videos indicate that the proposed algorithm ensures robust tracking and can be used effectively for real-time surveillance of highways.

MDL-CPI: Multi-view deep learning model for compound-protein interaction prediction.

Elucidating the mechanisms of Compound-Protein Interactions (CPIs) plays an essential role in drug discovery and development. Many computational efforts have been done to accelerate the development of this field. However, the current predictive performance is still not satisfactory, and existing methods consider only protein and compound features, ignoring their interactive information. In this study, we propose a multi-view deep learning method named MDL-CPI for CPI prediction. To sufficiently extract discriminative information, we introduce a hybrid architecture that leverages BERT (Bidirectional Encoder Representations from Transformers) and CNN (Convolutional Neural Network) to extract protein features from a sequential perspective, use the GNN (Graph Neural Networks) to extract compound features from a structural perspective, and generate a unified feature space by using AE2 (Autoencoder in Autoencoder Networks) network to learn the interactive information between BERT-CNN and Graph embeddings. Comparative results on benchmark datasets show that our proposed method exhibits better performance compared to existing CPI prediction methods, demonstrating the strong predictive ability of our model. Importantly, we demonstrate that the learned interactive information between compounds and proteins is critical to improve predictive performance. We release our source code and dataset at: https://github.com/Longwt123/MDL-CPI.

Few-shot class-incremental learning based on representation enhancement

Few-shot class-incremental learning (FSCIL) is crucial and practical for artificial intelligence in the real world, which learns novel classes incrementally from few samples without forgetting the previously learned classes. However, FSCIL confronts two significant challenges: “catastrophic forgetting” and “overfitting new.” We focus on convolutional neural network (CNN)-based FSCIL and propose a human cognition-inspired FSCIL method, in which the knowledge of novel classes is learned under the guidance of the previously learned knowledge. Specifically, we learn a discriminative and generalized CNN feature extractor from the base classes in the first task. We generate the representations of base and novel classes in unified feature space without training on novel classes, thus avoiding “forgetting old.” For the novel classes in long sequential tasks, beyond the representation generation, we enhance the representation by exploring the correlations with the previously learned classes to alleviate overfitting new and ensure that the novel classes adapt to the feature space. Experimental results show that our proposed method achieves very competitive results on MiniImageNet and CIFAR-100 datasets.

PIMA-CT: Physical Model-Aware Cyclic Simulation and Denoising for Ultra-Low-Dose CT Restoration.

A body of studies has proposed to obtain high-quality images from low-dose and noisy Computed Tomography (CT) scans for radiation reduction. However, these studies are designed for population-level data without considering the variation in CT devices and individuals, limiting the current approaches' performance, especially for ultra-low-dose CT imaging. Here, we proposed PIMA-CT, a physical anthropomorphic phantom model integrating an unsupervised learning framework, using a novel deep learning technique called Cyclic Simulation and Denoising (CSD), to address these limitations. We first acquired paired low-dose and standard-dose CT scans of the phantom and then developed two generative neural networks: noise simulator and denoiser. The simulator extracts real low-dose noise and tissue features from two separate image spaces (e.g., low-dose phantom model scans and standard-dose patient scans) into a unified feature space. Meanwhile, the denoiser provides feedback to the simulator on the quality of the generated noise. In this way, the simulator and denoiser cyclically interact to optimize network learning and ease the denoiser to simultaneously remove noise and restore tissue features. We thoroughly evaluate our method for removing both real low-dose noise and Gaussian simulated low-dose noise. The results show that CSD outperforms one of the state-of-the-art denoising algorithms without using any labeled data (actual patients' low-dose CT scans) nor simulated low-dose CT scans. This study may shed light on incorporating physical models in medical imaging, especially for ultra-low level dose CT scans restoration.

Semi-Supervised Ensemble Learning Framework for Accelerating Power System Transient Stability Knowledge Base Generation

Machine learning approaches have exhibited high potential in power system transient stability assessment (TSA), yet their initial preparation stages of stability knowledge base generation (SKBG) based on time-domain simulations often undergo high computational costs. In fact, how to ease the heavy computational burden of SKBG without sacrificing the reliability is still a significant challenge. To address this problem, this paper develops a semi-supervised ensemble learning (SSEL) framework for reliable SKBG acceleration, without additional need for computing hardware upgrading. In particular, it performs detailed simulations for a minority of cases and fast simulations for the majority ones to reduce the total computation time. Considering the absence of stability status information of those fast simulated cases, an SSEL scheme is systematically designed to reliably label their stability status. Before implementing SSEL, two concise feature descriptors are first introduced to efficiently extract transient features from multiplex system trajectories. Then, all the cases are characterized in a unified feature space, whereby a series of semi-supervised support vector machines are trained in randomly formed subspaces. Afterward, these single machines are systematically combined to derive an enhanced SSEL model, which is able to make reliable and robust labeling decisions. Further, a backtrace strategy is carefully devised for SSEL, so as to maintain the high reliability of SKBG. Test results on the IEEE 39-bus system and the realistic GD Power Grid in South China illustrate the excellent performances of the proposed framework on SKBG acceleration.

A Framework for Multivariate Analysis of Land Surface Dynamics and Driving Variables—A Case Study for Indo-Gangetic River Basins

The analysis of the Earth system and interactions among its spheres is increasingly important to improve the understanding of global environmental change. In this regard, Earth observation (EO) is a valuable tool for monitoring of long term changes over the land surface and its features. Although investigations commonly study environmental change by means of a single EO-based land surface variable, a joint exploitation of multivariate land surface variables covering several spheres is still rarely performed. In this regard, we present a novel methodological framework for both, the automated processing of multisource time series to generate a unified multivariate feature space, as well as the application of statistical time series analysis techniques to quantify land surface change and driving variables. In particular, we unify multivariate time series over the last two decades including vegetation greenness, surface water area, snow cover area, and climatic, as well as hydrological variables. Furthermore, the statistical time series analyses include quantification of trends, changes in seasonality, and evaluation of drivers using the recently proposed causal discovery algorithm Peter and Clark Momentary Conditional Independence (PCMCI). We demonstrate the functionality of our methodological framework using Indo-Gangetic river basins in South Asia as a case study. The time series analyses reveal increasing trends in vegetation greenness being largely dependent on water availability, decreasing trends in snow cover area being mostly negatively coupled to temperature, and trends of surface water area to be spatially heterogeneous and linked to various driving variables. Overall, the obtained results highlight the value and suitability of this methodological framework with respect to global climate change research, enabling multivariate time series preparation, derivation of detailed information on significant trends and seasonality, as well as detection of causal links with minimal user intervention. This study is the first to use multivariate time series including several EO-based variables to analyze land surface dynamics over the last two decades using the causal discovery algorithm PCMCI.

Visible-Infrared Person Re-Identification With Modality-Specific Memory Network.

Visible-infrared person re-identification (VI-ReID) is challenging due to the large modality discrepancy between visible and infrared images. Existing methods mainly focus on learning modality-shared representations by embedding images from different modalities into a common feature space, in which some discriminative modality information is discarded. Different from these methods, in this paper, we propose a novel Modality-Specific Memory Network (MSMNet) to complete the missing modality information and aggregate visible and infrared modality features into a unified feature space for the VI-ReID task. The proposed model enjoys several merits. First, it can exploit the missing modality information to alleviate the modality discrepancy when only the single-modality input is provided. To the best of our knowledge, this is the first work to exploit the missing modality information completion and alleviate the modality discrepancy with the memory network. Second, to guide the learning process of the memory network, we design three effective learning strategies, including feature consistency, memory representativeness and structural alignment. By incorporating these learning strategies in a unified model, the memory network can be well learned to propagate identity-related information between modalities and boost the VI-ReID performance. Extensive experimental results on two standard benchmarks (SYSU-MM01 and RegDB) demonstrate that the proposed MSMNet performs favorably against state-of-the-art methods.

A deep learning method for repurposing antiviral drugs against new viruses via multi-view nonnegative matrix factorization and its application to SARS-CoV-2.

The outbreak of COVID-19 caused by SARS-coronavirus (CoV)-2 has made millions of deaths since 2019. Although a variety of computational methods have been proposed to repurpose drugs for treating SARS-CoV-2 infections, it is still a challenging task for new viruses, as there are no verified virus-drug associations (VDAs) between them and existing drugs. To efficiently solve the cold-start problem posed by new viruses, a novel constrained multi-view nonnegative matrix factorization (CMNMF) model is designed by jointly utilizing multiple sources of biological information. With the CMNMF model, the similarities of drugs and viruses can be preserved from their own perspectives when they are projected onto a unified latent feature space. Based on the CMNMF model, we propose a deep learning method, namely VDA-DLCMNMF, for repurposing drugs against new viruses. VDA-DLCMNMF first initializes the node representations of drugs and viruses with their corresponding latent feature vectors to avoid a random initialization and then applies graph convolutional network to optimize their representations. Given an arbitrary drug, its probability of being associated with a new virus is computed according to their representations. To evaluate the performance of VDA-DLCMNMF, we have conducted a series of experiments on three VDA datasets created for SARS-CoV-2. Experimental results demonstrate that the promising prediction accuracy of VDA-DLCMNMF. Moreover, incorporating the CMNMF model into deep learning gains new insight into the drug repurposing for SARS-CoV-2, as the results of molecular docking experiments reveal that four antiviral drugs identified by VDA-DLCMNMF have the potential ability to treat SARS-CoV-2 infections.

Performance comparison of deep learning and machine learning methods in determining wetland water areas using EuroSAT dataset.

Wetlands are critical to the ecology because they maintain biodiversity and provide home for a variety of species. Researching, mapping, and conservation of wetlands is a challenging and time-consuming process. Because they produce temporal and geographical information, remote sensing and photogrammetric approaches are useful tools for analyzing and managing wetlands. In this study, the water areas of five different wetlands obtained with Sentinel-2 images in Turkey were classified. Although obtaining large amounts of high-dimensional dataset labeled for various land types is costly, it is a significant advantage to use it after model training in a wide range of applications. In this paper, the EuroSAT dataset was used in the validation process. Proposed deep learning-based 1D convolutional neural networks (CNN) and traditional machine learning methods (i.e., support vector machine, linear discriminant analysis, K-nearest neighborhood, canonical correlation forests, and AdaBoost.M1) were compared quantitatively (i.e., accuracy, recall, precision, specificity, F-score, and image quality assessment metrics) and qualitatively. Finally, pairwise comparison was made with chi-square-based McNemar's test. There is a statistical difference between 1D CNN and machine learning method (except the support vector machine vs linear discriminant analysis in Test 1 area). CNN models outperform machine learning algorithms in terms of non-linear function approximation and the ability to extract and articulate data features. Since 1D CNNs can process data in a highly complex and unique feature space, they are very successful in segmenting strongly related and highly correlated discrete signals. It also has advantages over machine learning methods for water body extraction in that it can be integrated with sophisticated image pre-processing and standardization tools, is less susceptible to low-level random noise, and provides shift in variations and contrast-invariant image local transforms.

Using physics to extend the range of machine learning models for an aerodynamic, hydraulic and combusting system: The toy model concept

Machine learning models used for energy conversion system optimization cannot extrapolate outside the bounds of training data and often produce physically unrealistic results when making predictions in regions of sparse training data. The toy model concept introduced in this work allows machine learning models to extrapolate to some extent and also reduces the possibility of physically unrealistic results. It uses physics to shrink the model input (feature space) of data-based models, so that extrapolations in the data-based feature space tend to become interpolations in the physics-based (toy variable) feature space. The physics-based model can be any model or experiment that can shrink the feature space without affecting interpolation and is termed a ‘toy model’ because it does not need to be accurate or make predictions of interest. The concept has been applied to model experimental data obtained from three complex systems: a. Aerodynamic forces on a spinning and vibrating baseball with inclined axis of rotation (toy model: CFD model), b. Hydraulic turbine efficiency (toy model: PIV images of flow through stationary blades), and c. Combustion generated engine emissions (toy model: system-level 1-D model). All extrapolations were converted into interpolations for the first two systems while a 75% conversion was achieved for the emission predictions. The engine toy model produced 736,281 possible feature spaces from which one unique feature space was chosen for every prediction based on agreement between different machine learning algorithms. It is shown that the ability of the toy variables to reorganize the data is important, while their accuracy is relatively unimportant. The toy model concept was demonstrated to work with neural networks and regression, and can be used to increase model robustness or reduce training data requirements.

PSSP-MVIRT: peptide secondary structure prediction based on a multi-view deep learning architecture.

The prediction of peptide secondary structures is fundamentally important to reveal the functional mechanisms of peptides with potential applications as therapeutic molecules. In this study, we propose a multi-view deep learning method named Peptide Secondary Structure Prediction based on Multi-View Information, Restriction and Transfer learning (PSSP-MVIRT) for peptide secondary structure prediction. To sufficiently exploit discriminative information, we introduce a multi-view fusion strategy to integrate different information from multiple perspectives, including sequential information, evolutionary information and hidden state information, respectively, and generate a unified feature space. Moreover, we construct a hybrid network architecture of Convolutional Neural Network and Bi-directional Gated Recurrent Unit to extract global and local features of peptides. Furthermore, we utilize transfer learning to effectively alleviate the lack of training samples (peptides with experimentally validated structures). Comparative results on independent tests demonstrate that our proposed method significantly outperforms state-of-the-art methods. In particular, our method exhibits better performance at the segment level, suggesting the strong ability of our model in capturing local discriminative information. The case study also shows that our PSSP-MVIRT achieves promising and robust performance in the prediction of new peptide secondary structures. Importantly, we establish a webserver to implement the proposed method, which is currently accessible via http://server.malab.cn/PSSP-MVIRT. We expect it can be a useful tool for the researchers of interest, facilitating the wide use of our method.

Deep Transfer Network With Multi-Kernel Dynamic Distribution Adaptation for Cross-Machine Fault Diagnosis

Recently, various deep learning models, which are mainly based on data-driven algorithms, have received more and more attention in the field of intelligent fault diagnosis and prognostics. However, there are two major assumptions accepted by default in the existing studies: 1) The training (source domain) and testing (target domain) data sets obey the same feature distribution; 2) Sufficient labeled data with fault information is available for model training. In real industrial scenarios, especially for different machines, these assumptions are mostly invalid, which makes it a huge challenge to build reliable diagnostic model. Motivated by transfer learning, we present a novel intelligent method named deep transfer network (DTN) with multi-kernel dynamic distribution adaptation (MDDA) to address the problem of cross-machine fault diagnosis. In the proposed approach, the DTN has wide first-layer convolutional kernel and several small convolutional layers, which is utilized to extract transferable features across different machines and suppress high frequency noise. Then, the MDDA method constructs a weighted mixed kernel function to map different transferable features to a unified feature space, and the relative importance of the marginal and conditional distributions are also evaluated dynamically. The proposed method is verified by three transfer learning tasks of bearings, in which the health states of wind turbine bearings in real scenario are identified by using diagnosis knowledge from two different bearings in laboratories. The results show that the proposed method can achieve higher diagnosis accuracy and better transfer performance even under different noisy environment conditions than many other state-of-the-art methods. The presented framework offers a promising approach for cross-machine fault diagnosis.

CrossATNet - a novel cross-attention based framework for sketch-based image retrieval

We propose a novel framework for cross-modal zero-shot learning (ZSL) in the context of sketch-based image retrieval (SBIR). Conventionally, the SBIR schema mainly considers simultaneous mappings among the two image views and the semantic side information. Therefore, it is desirable to consider fine-grained classes mainly in the sketch domain using highly discriminative and semantically rich feature space. However, the existing deep generative modeling based SBIR approaches majorly focus on bridging the gaps between the seen and unseen classes by generating pseudo-unseen-class samples. Besides, violating the ZSL protocol by not utilizing any unseen-class information during training, such techniques do not pay explicit attention to modeling the discriminative nature of the shared space. Also, we note that learning a unified feature space for both the multi-view visual data is a tedious task considering the significant domain difference between sketches and the color images. In this respect, as a remedy, we introduce a novel framework for zero-shot SBIR. While we define a cross-modal triplet loss to ensure the discriminative nature of the shared space, an innovative cross-modal attention learning strategy is also proposed to guide feature extraction from the image domain exploiting information from the respective sketch counterpart. In order to preserve the semantic consistency of the shared space, we consider a graph CNN based module which propagates the semantic class topology to the shared space. To ensure an improved response time during inference, we further explore the possibility of representing the shared space in terms of hash-codes. Experimental results obtained on the benchmark TU-Berlin and the Sketchy datasets confirm the superiority of CrossATNet in yielding the state-of-the-art results.

Local ranking and global fusion for personalized recommendation

Most of top-n personalized recommendation algorithms perform a kind of global learning-to-rank based on the original user–item rating matrix, yet with an implicit presumption of a unified feature space for all users (all items). However, we argue that such a presumption is a limiting factor for better recommendation, because features learned in a global space could be too coarse-grained. As such, ranking a recommendation list against a unified space might be dominated by only a few significant aspects of users’ interests.In this paper, we propose a novel local ranking and global fusion (LRGF) framework for personalized recommendation, which allows using multiple fine-grained feature spaces for learning-to-rank. In LRGF, we first adopt two-stage random walk on the established user graph and item graph for selecting those mostly related users and items, so as to form multiple local user–item groups. Next, we propose a pairwise gap-aware Bayesian personalized ranking to obtain local ranking lists for each local user–item group. Finally, a global decision fusion is proposed to obtain the final top-n recommendation list for each user only from her local ranking lists. Experiments on the three real-world datasets show that the proposed LRGF outperforms the state-of-the-art schemes in terms of higher mean average precision and normalized discounted cumulative gain.

Bearing State Recognition Method Based on Transfer Learning Under Different Working Conditions.

Bearing state recognition, especially under variable working conditions, has the problems of low reusability of monitoring data, low state recognition accuracy and low generalization ability of the model. The feature-based transfer learning method can solve the above problems, but it needs to rely on signal processing knowledge and expert diagnosis experience to obtain the cross-characteristics of different working conditions data in advance. Therefore, this paper proposes an improved balanced distribution adaptation (BDA), named multi-core balanced distribution adaptation (MBDA). This method constructs a weighted mixed kernel function to map different working conditions data to a unified feature space. It does not need to obtain the cross-characteristics of different working conditions data in advance, which simplifies the data processing and meet end-to-end state recognition in practical applications. At the same time, MBDA adopts the A–Distance algorithm to estimate the balance factor of the distribution and the balance factor of the kernel function, which not only effectively reduces the distribution difference between different working conditions data, but also improves efficiency. Further, feature self-learning and rolling bearing state recognition are realized by the stacked autoencoder (SAE) neural network with classification function. The experimental results show that compared with other algorithms, the proposed method effectively improves the transfer learning performance and can accurately identify the bearing state under different working conditions.

Reconstructed similarity for faster GANs-based word translation to mitigate hubness

In machine word translation, Nearest Neighbor (NN) retrieval is able to search the best-k translation candidates as linguistic labels of a source query from a unified multilingual semantic feature space. However, NN is polluted by hubs in the high-dimensional feature space. Many proposed approaches remove hubs in the list of translation candidates to relieve this problem. But those approaches to eliminating hubs are flawed because they also have corresponding translations. To address this issue, we propose a novel Reconstructed Similarity (RS) retrieval for the neural machine word translation model to mitigate the hubness problem regardless of whether it is a hub. Different from previous work, RS reduces the impact of hubness pollution in dense and high-dimensional space and allows the hubs to have the same probability as the target candidates without being inappropriately excluded. In addition, RS improves the quality of bilingual dictionaries by measuring the bilateral similarity of the bilingual and monolingual distance of each of the source query embeddings. Additionally, to model the unsupervised machine word translation, we introduce Generative Adversarial Nets (GANs) to map the source and target word distribution into a shared semantic space. We also construct a tiny GAN topology for neural machine word translation, which is at least 52 × faster than previous GAN-based models. To further align cross-lingual embedding distributions, we provide orthogonal Procrustes mapping, global-awareness of the transformation matrix and rescaling of the target embeddings as flexible and optional multirefinements. The results show that our model outperforms the state-of-the-art by nearly 4% in distant languages such as English to Finnish. Compared with a precision@1 of 47.00% from English to Finnish, our model obtains a precision@1 of 47.53% and achieves state-of-the-art results in a fully unsupervised form. Moreover, our model achieves competitive results in the shortest time among GAN-based models, which easily trade off between speed and accuracy.

Generative adversarial fusion network for class imbalance credit scoring

Credit scoring on class imbalance data, where the class of defaulters is insufficiently represented compared with the class of non-defaulters, is an important but challenging task. In this paper, we propose an imbalanced generative adversarial fusion network (IGAFN) to cope with the class imbalance credit scoring based on multi-source heterogeneous credit data. Concretely, we design a fusion module to integrate the heterogeneous credit data from multiple sources into a unified latent feature space. A generative adversarial network-based balance module is then designed to generate latent representations of new samples for the minority class of the imbalanced datasets. The performance of IGAFN is compared against multiple conventional machine learning and deep learning algorithms. Extensive experiments show that the proposed IGAFN exhibits significantly better performance than the compared methods on two real-life datasets.