Distribution Divergence Research Articles

Divergence-Guided Simultaneous Speech Translation

To achieve high-quality translation with low latency, a Simultaneous Speech Translation (SimulST) system relies on a policy module to decide whether to translate immediately or wait for additional streaming input, along with a translation model capable of effectively handling partial speech input. Prior research has tackled these components separately, either using ``wait-k'' policies based on fixed-length segments or detected word boundaries, or dynamic policies based on different strategies (e.g., meaningful units), while employing offline models for prefix-to-prefix translation. In this paper, we propose Divergence-Guided Simultaneous Speech Translation (DiG-SST), a tightly integrated approach focusing on both translation quality and latency for streaming input. Specifically, we introduce a simple yet effective prefix-based strategy for training translation models with partial speech input, and develop an adaptive policy that makes read/write decisions for the translation model based on the expected divergence in translation distributions resulting from future input. Our experiments on multiple translation directions of the MuST-C benchmark demonstrate that our approach achieves a better trade-off between translation quality and latency compared to existing methods.

AMD: Autoregressive Motion Diffusion

Human motion generation aims to produce plausible human motion sequences according to various conditional inputs, such as text or audio. Despite the feasibility of existing methods in generating motion based on short prompts and simple motion patterns, they encounter difficulties when dealing with long prompts or complex motions. The challenges are two-fold: 1) the scarcity of human motion-captured data for long prompts and complex motions. 2) the high diversity of human motions in the temporal domain and the substantial divergence of distributions from conditional modalities, leading to a many-to-many mapping problem when generating motion with complex and long texts. In this work, we address these gaps by 1) elaborating the first dataset pairing long textual descriptions and 3D complex motions (HumanLong3D), and 2) proposing an autoregressive motion diffusion model (AMD). Specifically, AMD integrates the text prompt at the current timestep with the text prompt and action sequences at the previous timestep as conditional information to predict the current action sequences in an iterative manner. Furthermore, we present its generalization for X-to-Motion with “No Modality Left Behind”, enabling for the first time the generation of high-definition and high-fidelity human motions based on user-defined modality input.

A dual‐channel transferable RUL prediction method integrated with Bayesian deep learning and domain adaptation for rolling bearings

AbstractMany deep learning methods typically assume that the marginal probability distribution between the training and testing bearing data is similar or the same. However, the probability distribution of rolling bearings may deviate significantly under diverse working conditions. To address the above limitations, a novel transferable remaining useful life (RUL) prediction method integrated with Bayesian deep learning and unsupervised domain adaptation (DA) is proposed. First, the signal alignment is executed on the data after the first prediction time to maintain the same granularity and scale across both source and target domains. Second, the multi‐domain features are extracted and sent into the dual‐channel Transformer network (DCTN) incorporating the convolutional block attention module (CBAM) to adequately exploit the abundant degradation information. Then, the DA module is incorporated into the model to mitigate the distribution discrepancies of the extracted high‐level merged features between the source and target domains. Finally, by applying the variational inference method, the DCTN‐CBAM is extended to the Bayesian deep neural network, and the RUL prediction and its corresponding confidence intervals can be conveniently derived. In addition, the generalization capability and effectiveness are validated through six bidirectional transfer RUL prediction tasks across two rolling bearing datasets. The experimental results demonstrate that it could provide a more reliable RUL prediction and efficiently account for the prediction uncertainty.

TESTING THE DISTRIBUTIONAL DIVERGENCE IN THE PILLARS OF REGIONAL DEVELOPMENT IN MOROCCO

This article examines the distributional divergence of the pillars of regional development in Morocco using a new hierarchical clustering algorithm from the unsupervised machine learning literature. The study utilizes a unique dataset of regional indicators to identify the major clusters of development and differences in regional performance in terms of economic transformation, social inclusion, and environmental sustainability. Results from the hierarchical clustering algorithm show that the Moroccan regions are highly differentiated and the clusters identified do not necessarily coincide with the traditional administrative divisions. This suggests that the policies implemented in Morocco have not been effective in achieving balanced development across all regions. The findings of this study provide important insights into the challenges of regional development in Morocco and the potential of using machine learning algorithms to better identify and address regional disparities.

MNEMONIC: Multikernel contrastive domain adaptation for time-series classification

Time-Series Classification (TSC) has gained substantial importance in applications such as healthcare, finance, manufacturing, and human activity recognition. Training and evaluating a model on data with diverse distributions pose challenges across domains, but Unsupervised Domain Adaptation (UDA) proves efficient in mitigating these distribution discrepancies. Current TSC-UDA methods exploit statistical, adversarial, or contrastive learning strategies. Even though contrastive learning is an emerging UDA, relying solely on it proves inadequate in addressing the non-linear characteristics of the non-stationary multivariate temporal data. Kernelization enables better capturing of these non-linear relationships among the samples. Hence, an optimal kernel is derived by applying multiple kernel variants to samples and the maximum mean discrepancy metric is computed. The augmented samples in contrastive learning can result in overfitting, hence L1norm regularization is adopted to eliminate it. Therefore, we introduce a comprehensive approach Multikernel coNtrastivE doMain adaptatiON for tIme-series Classification (MNEMONIC) that embraces statistical feature alignment, contrastive learning, and regularization. Consequently, the model learns the inherent relationship among non-monotonic temporal data points and aligns the domains into a shared subspace, reducing the distribution differences. Extensive experimental evaluation is performed and the results are compared to the state-of-the-art with a macro-F1 score of 94.36%, 86.19%, 70.08%, and 63.07% for UCIHAR, HHAR, WISDM, and SSC datasets respectively.

A novel adaptive deep transfer learning method towards thermal error modeling of electric spindles under variable conditions

Thermal error modeling (TEM) plays a vital role in maintaining the machining accuracy of electric spindles. Recently, deep learning (DL) techniques have obtained promising achievements in this area. However, DL techniques have certain limitations. The data acquired from variable working conditions present large distribution discrepancies, the DL-based model established on one working condition fails to obtain satisfactory prediction accuracy in another working condition. Moreover, existing studies focus solely on using temperature features to build prediction models, neglecting the full use of multi-sensory information. To address these issues, this paper proposes a novel adaptive deep transfer learning method towards TEM of electric spindles, which takes full advantage of the temperature, current, and power sensory information. Firstly, finite element simulation is employed to analyze the thermal characteristics of the electric spindle and determine the locations for temperature measurement points. Then, the convolutional long short-term memory network (C-LSTMN) is constructed, where the spatial features from multi-sensory information between inputs and prediction patterns are captured by convolutional layers, these features are further processed by long short-term memory network (LSTMN) to extract temporal features. Subsequently, the multi-kernel joint maximum mean discrepancy (MK-JMMD) measure is developed to minimize the distribution discrepancies between the source and target domains, thus the prediction model initially established on the source domain can be adaptive to effectively predict on the target domain. Finally, with unavailable thermal error data in the target domain, the proposed method is validated through 12 transfer tasks using datasets from four working conditions with other comparison methods. The results demonstrate that the proposed method overcomes the challenges of unavailable labeled thermal error samples and outperforms advanced methods.

On the Size Discrepancies between Datasets from China Meteorological Administration and Joint Typhoon Warning Center for the Northwestern Pacific Tropical Cyclones

This study analyzes the Northwestern Pacific tropical cyclone (TC) size difference between the China Meteorological Administration (CMA) dataset and the Joint Typhoon Warning Center (JTWC) dataset. The TC size is defined by the near-surface 34-knot wind radius (R34). Although there is a high correlation (correlation coefficient of 0.71) between CMA and JTWC R34 values, significant discrepancies are still found between them. The JTWC tends to report larger R34 values than the CMA for large-sized TCs, while the trend is reversed for compact TCs. Despite spatial distribution discrepancies, both datasets exhibit significant similarity (spatial correlation coefficient of 0.61), particularly in latitudinal distribution; higher R34 values are observed near 25° N. An investigation of key parameters affecting R34 estimations shows that the discrepancies in R34 values between the two agencies’ estimates of TC size are primarily influenced by the size itself and latitude. There is a high correlation between R34 difference and R34 values, with a high correlation of up to 0.58 with the JTWC’s R34 values. There is also a significant correlation between R34 difference and latitude, with a correlation coefficient of 0.26 in both the CMA and JTWC datasets. Case studies of Typhoons “Danas” and “Maysak” confirm distinct characteristics in R34 estimations during different development stages, with the JTWC capturing TC intensification better, while the CMA underestimates TC size during rapid growth phases. During the weakening stage of the TC, both agencies accurately estimate the R34 values. These findings contribute valuable insights into the discrepancies and characteristics of R34 datasets, informing the selection and utilization of data for typhoon research and forecasting.

Transfer Learning for Bearing Fault Diagnosis based on Graph Neural Network with Dilated KNN and Adversarial Discriminative Domain Adaptation

Graph neural networks (GNNs) have emerged as a forefront in deep learning, notably influencing research in mechanical fault diagnosis. Transfer learning, particularly through domain adaptation (DA) techniques, has found application in machinery fault diagnosis by training models under one working condition and deploying them under another. While efforts have been made to integrate GNNs with DA techniques to alleviate data distribution discrepancies by investigating the inter-sample relationships, challenges persist: reliance on K-nearest neighbor (KNN) for graph generation emphasizes close relationships, neglecting distant ones; batch processing limits real-time fault diagnosis; and transfer between different-sized bearings is nearly unexplored. To address these limitations, a novel framework for GNN-based domain adaptation in machinery fault diagnosis is proposed. Initially, a convolutional neural network extracts node embeddings from the continuous wavelet transform graph of raw vibration signals. Subsequently, a graph generation layer based on dilated KNN captures both close and distant sample relationships, addressing the long-range dependency issue. Two GNN blocks are then applied for inter-sample relationships investigation and further feature extraction with the outputs directed to a linear classifier during source domain pretraining. Following pretraining, adversarial discriminative domain adaptation is leveraged to mitigate domain distribution discrepancies. Additionally, a novel graph construction method that combines existing training samples with a new single sample is proposed, enabling fault prediction with single instances for real-time online fault diagnosis. Evaluation on datasets with varying working conditions and bearings of different sizes demonstrates the superior performance of our method to other comparison methods.

Domain adaptive remote sensing image semantic segmentation with prototype guidance

Current unsupervised domain adaptation (UDA) techniques in semantic segmentation effectively decrease the domain discrepancy between the labeled source domain and unlabeled target domain, thereby enhancing the model’s pixel-wise discriminative capability for target domain data. However, in remote sensing images (RSIs), our study uncovers that these approaches perform poorly in the presence of class distribution inconsistencies between the source and target domains. In this work, we propose a one-stage mean teacher framework with a novel auxiliary prototype classifier, named MTA, to address this issue. Specifically, the teacher model assigns pseudo labels at pixel level for target samples and captures knowledge from the student model via exponential moving average (EMA). With labeled source samples and target samples that have pseudo labels, the student model can alleviate the divergence between the source and target domains. In addition, the auxiliary prototype classifier (APC) reduces the performance degradation in the parametric softmax classifier of the student model caused by class distribution divergence. We also propose a prototype computation scheme to obtain each class prototype in the APC. Specifically, we build a memory bank for each class of the two domains to store feature embeddings dynamically. Then, we compute the class prototype by applying the clustering algorithm on memory banks corresponding to the class. Meanwhile, the APC reduces the intra-class domain discrepancy by optimizing the cross-entropy loss, which brings each class feature distribution of the two domains closer to the class prototype. The experimental results on RSIs UDA semantic segmentation tasks show the superiority of our approach over comparative methods.

An Approach to Incorporating Implicit Knowledge in Object Detection Models

Current object detection methods typically focus on addressing the distribution discrepancies between source and target domains. However, solely concentrating on this aspect may lead to overlooking the inherent limitations of the samples themselves. This study proposes a method to integrate implicit knowledge into object detection models, aiming to enhance the models’ effectiveness in identifying target features within images. We analyze the sources of information loss in object detection models, treating this loss as a form of implicit knowledge and modeling it in the form of dictionaries. We explore potentially effective ways of integrating latent knowledge into the models and then apply it to object detection models. The models demonstrate a 1% and 0.8% improvement in mean average precision(mAP) in the UA-DETRAC and KITTI datasets, respectively. The results indicate that the proposed method can effectively enhance the relevant metrics of object detection models without significantly increasing the parameter or computational overhead. By excavating and utilizing implicit knowledge, we enhance the performance and efficiency of the models, offering new perspectives and methods for addressing challenges in practical applications.

Small is Beautiful: Compressing Deep Neural Networks for Partial Domain Adaptation.

Domain adaptation is a promising way to ease the costly data labeling process in the era of deep learning (DL). A practical situation is partial domain adaptation (PDA), where the label space of the target domain is a subset of that in the source domain. Although existing methods yield appealing performance in PDA tasks, it is highly presumable that computation overhead exists in deep PDA models since the target is only a subtask of the original problem. In this work, PDA and model compression are seamlessly integrated into a unified training process. The cross-domain distribution divergence is reduced by minimizing a soft-weighted maximum mean discrepancy (SWMMD), which is differentiable and functions as regularization during network training. We use gradient statistics to compress the overparameterized model to identify and prune redundant channels based on the corresponding scaling factors in batch normalization (BN) layers. The experimental results demonstrate that our method can achieve comparable classification performance to state-of-the-art methods on various PDA tasks, with a significant reduction in model size and computation overhead.

Multiscale Change Detection Domain Adaptation Model Based on Illumination–Reflection Decoupling

In the change detection (CD) task, the substantial variation in feature distributions across different CD datasets significantly limits the reusability of supervised CD models. To alleviate this problem, we propose an illumination–reflection decoupled change detection multi-scale unsupervised domain adaptation model, referred to as IRD-CD-UDA. IRD-CD-UDA maintains its performance on the original dataset (source domain) and improves its performance on unlabeled datasets (target domain) through a novel CD-UDA structure and methodology. IRD-CD-UDA synergizes mid-level global feature marginal distribution domain alignment, classifier layer feature conditional distribution domain alignment, and an easy-to-hard sample selection strategy to increase the generalization performance of CD models on cross-domain datasets. Extensive experiments conducted on the LEVIR, SYSU, and GZ optical remote sensing image datasets demonstrate that the IRD-CD-UDA model effectively mitigates feature distribution discrepancies between source and target CD data, thereby achieving optimal recognition performance on unlabeled target domain datasets.

Style-Guided Adversarial Teacher for Cross-Domain Object Detection

The teacher–student framework is widely employed for cross-domain object detection. However, it suffers from two problems. One is that large distribution discrepancies will cause critical performance drops. The other is that the samples that deviate from the overall distributions of both domains will greatly mislead the model. To solve these problems, we propose a style-guided adversarial teacher (SGAT) method for domain adaptation. Specifically, on the domain level, we generate target-like images based on source images to effectively narrow the gaps between domains. On the sample level, we denoise samples by estimating the probability density ratio of the ‘target-style’ and target distributions, which could filter out the unrelated samples and highlight the related ones. In this way, we could guarantee reliable samples. With these reliable samples, we learn the domain-invariant features through teacher–student mutual learning and adversarial learning. Extensive experiments verify the effectiveness of our method. In particular, we achieve 52.9% mAP on Clipart1k and 42.7% on Comic2k, which are 6.4% and 5.0% higher than the compared baselines.

Unraveling the spatial-temporal distribution patterns of soil abundant and rare bacterial communities in China's subtropical mountain forest.

The pivotal roles of both abundant and rare bacteria in ecosystem function are widely acknowledged. Despite this, the diversity elevational patterns of these two bacterial taxa in different seasons and influencing factors remains underexplored, especially in the case of rare bacteria. Here, a metabarcoding approach was employed to investigate elevational patterns of these two bacterial communities in different seasons and tested the roles of soil physico-chemical properties in structuring these abundant and rare bacterial community. Our findings revealed that variation in elevation and season exerted notably effects on the rare bacterial diversity. Despite the reactions of abundant and rare communities to the elevational gradient exhibited similarities during both summer and winter, distinct elevational patterns were observed in their respective diversity. Specifically, abundant bacterial diversity exhibited a roughly U-shaped pattern along the elevation gradient, while rare bacterial diversity increased with the elevational gradient. Soil moisture and N:P were the dominant factor leading to the pronounced divergence in elevational distributions in summer. Soil temperature and pH were the key factors in winter. The network analysis revealed the bacteria are better able to adapt to environmental fluctuations during the summer season. Additionally, compared to abundant bacteria, the taxonomy of rare bacteria displayed a higher degree of complexity. Our discovery contributes to advancing our comprehension of intricate dynamic diversity patterns in abundant and rare bacteria in the context of environmental gradients and seasonal fluctuations.

Elucidating Optimal Exhaust Manifold Divergence and Temperature Distribution in Improving Low-End Engine Speed Performance

The exhaust manifold plays a crucial role in optimizing the performance of Spark-Ignition (SI) engines by effectively expelling combustion products. This study focuses on the optimization of the exhaust manifold design and temperature distribution in a 115-cc single-cylinder SI engine. The objective is to investigate the association between the design characteristics of the exhaust manifold, particularly its divergence, and the engine's performance in terms of brake power. Using computer-aided design (CAD), a three-dimensional model of the exhaust manifold with reduced diameter was developed. The optimized design aimed to enhance the engine's overall performance by achieving lower temperatures, particularly at low-end speeds. Subsequently, a 1D engine study was conducted to evaluate the performance of the engine with the optimized exhaust manifold design and validate the improved temperature distribution. The results demonstrate that the optimized exhaust manifold design leads to higher brake power while maintaining lower temperatures, especially at low-end speeds. This highlights the importance of exhaust manifold optimization and temperature distribution in maximizing the efficiency of the selected SI engine. To further enhance the engine's performance, future research should focus on identifying the most appropriate value for the exhaust manifold's divergence. This will contribute to the ongoing development of more efficient and high-performance SI engines. The findings of this study provide valuable insights into the optimization of exhaust manifold design and temperature distribution for improved engine performance in the context of the tested engine.

Transfer learning rolling bearing fault diagnosis model based on deep feature decomposition and class-level alignment

Research on transfer learning in rolling bearing fault diagnosis can help overcome challenges such as different data distributions and limited fault samples. However, most existing methods still struggle to address the zero-shot cross-domain problem within the same equipment and the few-shot cross-machine problem. In response to these challenges, this paper introduces a transfer learning rolling bearing fault diagnosis model based on deep feature decomposition and class-level alignment. The model consists of two stages. In the first stage, the original vibration signals undergo continuous wavelet transform to obtain time–frequency diagram. Subsequently, a convolutional neural network extracts features from the diagram. The obtained deep features are decomposed into four types: uncertain features, domain-shared features, domain-specific features and category features. Multiple loss functions are then employed to remove extraneous features beyond the category features. In the second stage, category features are further extracted, and convolutional block attention module is introduced to further reduce the potential interference of unexcluded irrelevant information within the category features with classification results. Simultaneously applying a class-level alignment strategy effectively alleviates inter-domain class distribution discrepancies. Experimental validation was conducted on three distinct datasets, revealing a significant improvement in the classification performance of the proposed method over alternative methods. Furthermore, the model demonstrated robustness and noise resistance.

FedPT-V2G: Security enhanced federated transformer learning for real-time V2G dispatch with non-IID data

The rising popularity of electric vehicles (EVs) underscores the potential of vehicle-to-grid (V2G) technology to contribute to load peak-shaving, valley-filling, and photovoltaic (PV) self-consumption. Effective V2G control strategies can be obtained by data-driven techniques, which is able to leverage historical and current data to inform future decision-making amidst uncertainties. However, the centralized collection and sharing of data among charging stations face challenges due to data asset concerns. Furthermore, even if data sharing hurdles are overcome, the non-independent and non-identically distributed (Non-IID) nature of data across charging stations can still negatively impact performance. In this study, we introduce FedPT-V2G, a security-enhanced federated transformer learning approach for real-time V2G dispatch that addresses Non-IID data. We employ deep learning models trained on historical and current data to enable real-time decision-making, facilitating both load shifting and PV self-consumption. Additionally, we utilize federated learning to jointly train a global model across all charging stations without collecting or sharing any local private data. We pioneer the application of the Proximal algorithm and Transformer model to tackle data distribution discrepancies within the V2G scheduling prediction task. The Proximal algorithm employs regularization techniques to align local models at each charging station more closely with the global model during updates. Concurrently, the multi-head attention mechanism within the Transformer model allows learned feature vectors to diverge, enabling better exploitation of variations across the entire feature space. Finally, we validate the proposed FedPT-V2G approach through extensive numerical simulations, demonstrating comparable performance to centralized learning on both balanced (98.93% vs 98.65%) and imbalanced (92.15% vs 92.20% in label skew) datasets.

Low thermal SRR metamaterial design with multi-layered structured for terahertz frequency applications

The goal of this study is to examine the effects of split-ring resonators (SRR) with compact square shapes at terahertz frequencies. Therefore, a unique and miniaturised SRR metamaterial was constructed in this work. Every simulation was performed by using the Computer Simulation Technology software to explore the radar cross-section (RCS), reflection and transmission coefficients. Furthermore, MATLAB software was adopted to calculate the characteristics of an effective medium for the recommended SRR metamaterial. The initial metamaterial structure consists of a single layer that exhibits lower peak points with unsatisfactory magnitude values. Therefore, by adopting the trial and error technique, a few parametric studies were conducted to select the optimised metamaterial. For instance, various square-shaped design analyses, investigation of different substrate materials with a dielectric constant between 3.5 and 11.9, multi-layer structures, and larger-scale analyses were explored in this work. According to the observation, the quintuple layer SRR design was adopted since it possesses multiple resonances with reasonable magnitude values at 2.58, 3.78, 7.54, 7.77, 8.21, 8.5, and 8.88 THz. Moreover, the proposed quintuple-layer metamaterial helps to reduce the RCS values from −128.5 to −90.93 dBm2 at 2.58 THz when adopting an array cell structure. Moreover, this work also examines and analyses the thermal distribution of the proposed quintuple-layer metamaterial at all resonance frequencies. The observation revealed only small temperature distribution discrepancies occurred on the design structure compared to all peak points. In a nutshell, the proposed quintuple-layer SRR metamaterial achieved the purpose of this research and is practically applicable for the terahertz frequency.

Alignment-Based Adversarial Training (ABAT) for Improving the Robustness and Accuracy of EEG-Based BCIs.

Machine learning has achieved great success in electroencephalogram (EEG) based brain-computer interfaces (BCIs). Most existing BCI studies focused on improving the decoding accuracy, with only a few considering the adversarial security. Although many adversarial defense approaches have been proposed in other application domains such as computer vision, previous research showed that their direct extensions to BCIs degrade the classification accuracy on benign samples. This phenomenon greatly affects the applicability of adversarial defense approaches to EEG-based BCIs. To mitigate this problem, we propose alignment-based adversarial training (ABAT), which performs EEG data alignment before adversarial training. Data alignment aligns EEG trials from different domains to reduce their distribution discrepancies, and adversarial training further robustifies the classification boundary. The integration of data alignment and adversarial training can make the trained EEG classifiers simultaneously more accurate and more robust. Experiments on five EEG datasets from two different BCI paradigms (motor imagery classification, and event related potential recognition), three convolutional neural network classifiers (EEGNet, ShallowCNN and DeepCNN) and three different experimental settings (offline within-subject cross-block/-session classification, online cross-session classification, and pre-trained classifiers) demonstrated its effectiveness. It is very intriguing that adversarial attacks, which are usually used to damage BCI systems, can be used in ABAT to simultaneously improve the model accuracy and robustness.

Select, Purify, and Exchange: A Multisource Unsupervised Domain Adaptation Method for Building Extraction.

Accurately extracting buildings from aerial images has essential research significance for timely understanding human intervention on the land. The distribution discrepancies between diversified unlabeled remote sensing images (changes in imaging sensor, location, and environment) and labeled historical images significantly degrade the generalization performance of deep learning algorithms. Unsupervised domain adaptation (UDA) algorithms have recently been proposed to eliminate the distribution discrepancies without re-annotating training data for new domains. Nevertheless, due to the limited information provided by a single-source domain, single-source UDA (SSUDA) is not an optimal choice when multitemporal and multiregion remote sensing images are available. We propose a multisource UDA (MSUDA) framework SPENet for building extraction, aiming at selecting, purifying, and exchanging information from multisource domains to better adapt the model to the target domain. Specifically, the framework effectively utilizes richer knowledge by extracting target-relevant information from multiple-source domains, purifying target domain information with low-level features of buildings, and exchanging target domain information in an interactive learning manner. Extensive experiments and ablation studies constructed on 12 city datasets prove the effectiveness of our method against existing state-of-the-art methods, e.g., our method achieves 59.1% intersection over union (IoU) on Austin and Kitsap → Potsdam, which surpasses the target domain supervised method by 2.2% . The code is available at https://github.com/QZangXDU/SPENet.