Beyond Sentence-Level End-to-End Speech Translation: Context Helps

To improve human-machine cooperation and enhance the accuracy of gait recognition in wearable lower limb exoskeletons, an enhancement method of gait recognition based on adaptive feature selection and an optimized machine learning algorithm was proposed. In this study, surface electromyography (sEMG) signals of rectus femoris, medialis femoris, lateralis femoris, semitendinosus and biceps femoris were recorded during level-ground walking. Then, time domain, frequency domain, time-frequency domain and nonlinear features were extracted. The integrated value of electromyography, variance, root mean square and wavelength were selected as the time domain features, and the mean power frequency and median frequency were selected as the frequency domain features. Wavelet packet energy was selected as the time-frequency domain feature. Nonlinear features including approximate entropy, sample entropy and fuzzy entropy of sEMG were extracted. To reduce feature dimension and improve the calculation efficiency, adaptive feature selection was performed by particle swarm optimization combined with sigmoid function. Then, the feature matrix was determined as the input for a random forest classifier to recognize different gait phases. An adaptive optimization mechanism based on Bayesian optimization was applied to random forest. Compared with random forest, the overall performance of the optimized model was improved. Its accuracy was increased by 3.57%. The feature selection and adaptive optimization mechanisms in gait recognition not only enhance the accuracy of random forest algorithms applied to sEMG for gait prediction, but also facilitate the flexibility and consistency required for lower limb exoskeleton gait control.

The examination of sleep patterns in newborns, particularly premature infants, is crucial for understanding neonatal development. This study presents an automated multi-sleep state classification approach for infants in neonatal intensive care units (NICU) using multiperspective feature extraction methodologies and machine learning to assess their neurological and physical development. The datasets for this study were collected from Children's Hospital Fudan University, Shanghai and consist of electroencephalography (EEG) recordings from two datasets, one comprising 64 neonates and the other 19 neonates. The proposed study involves six major phases: data collection, data annotation, preprocessing, multi-perspective feature extraction, adaptive feature selection, and classification. During the preprocessing phase, noise reduction is achieved using the multi-scale principal component analysis (MSPCA) method. From each epoch of eight EEG channels, a diverse ensemble of 1,976 features is extracted. This extraction employs a combination of stationary wavelet transform (SWT), flexible analytical wavelet transform (FAWT), spectral features based on $\alpha$, $\beta$, $\theta$, and $\delta$ brain waves, and temporal features refined through adaptive feature algorithm. In terms of performance, the proposed approach demonstrates significant improvements over existing studies. Using a single EEG channel, the model achieves accuracy of 81.45% and a Kappa score of 71.75%. With four channels, these metrics increase to 83.71% accuracy and a 74.04% Kappa score. Furthermore, utilizing all eight channels, the mean accuracy reaches to 85.62%, and the Kappa score rises to 76.30%. To evaluate the model's effectiveness, a leave-one-subject-out cross-validation method is employed. This thorough analysis validates the reliability of the classification approach. This makes it a promising method for monitoring and assessing sleep patterns in neonates.

Conventional computed tomography (CT) images often suffer from blurred edges and unclear details. Image super‐resolution methods can significantly enhance CT image quality, thereby improving diagnostic accuracy. To better extract detailed features and enhance the cascading effects of different feature levels, we propose a novel medical image super‐resolution algorithm that integrates discrete wavelet transform and multi‐scale adaptive feature selection. Our approach uses both the low‐resolution image and its high‐frequency component from the frequency domain as network inputs, with the high‐frequency component providing learning supervision, which enhances detail fidelity in reconstruction. Additionally, we introduce a multi‐scale adaptive feature selection module to learn from different layers of CT images and their inter‐layer correlations. Finally, the pixel information is efficiently integrated by a coordinate attention mechanism incorporating the concept of squeeze excitation. Experimental results show that our method outperforms state‐of‐the‐art methods, achieving superior reconstruction at scale factors of 2, 4, and 8, especially at scale factor 8, where it surpasses others by 1.12 in PSNR, 0.0145 in SSIM, and 0.0038 in LPIPS. Visually, our method also delivers more accurate details and better perceptual quality.

Tire defects are crucial for safe driving. Specialized experts or expensive tools such as stereo depth cameras and depth gages are usually used to investigate these defects. In image processing, feature extraction, reduction, and classification are presented as three challenging and symmetric ways to affect the performance of machine learning models. This paper proposes a hybrid system for cracked tire detection based on the adaptive selection of correlation features and deep belief neural networks. The proposed system has three steps: feature extraction, selection, and classification. First, the oriented gradient histogram extracts features from the tire images. Second, the proposed adaptive correlation feature selection selects important features with a threshold value adapted to the nature of the images. The last step of the system is to predict the image category based on the deep belief neural networks technique. The proposed model is tested and evaluated using real images of cracked and normal tires. The experimental results show that the proposed solution performs better than the current studies in effectively classifying tire defect images. The proposed hybrid cracked tire detection system based on adaptive correlation feature selection and Deep Belief Neural Networks’ performance provided better classification accuracy (88.90%) than that of Belief Neural Networks (81.6%) and Convolution Neural Networks (85.59%).

Electrocardiogram (ECG) analysis is one of the most important approaches to cardiac arrhythmia detection. This paper proposed an ECG analysis approach with adaptive feature selection and support vector machines (SVMs). Many wavelet transform-based coefficients are used as candidates, but only a few coefficients are selected for classification problem of each class pair. In addition, the several variation classes are partitioned into two or more subclasses to improve the training efficiency of SVMs. The experimental results show that the proposed ECG analysis approach can obtain high recognition rate and reliable results

A semantic parsing model for English video scenes based on sentence semantics and adaptive feature selection is proposed, which dynamically optimises visual features through text guidance and achieves cross modal semantic alignment.The experiment shows that the model performs well on multiple datasets: the average intersection to union ratio on ActiveNet reaches 91, exceeding other models by more than 11 points; Only 60 rounds of training on InternVid achieved an average intersection to union ratio of 90.The model parameter size is 185M, and the inference speed on RTX 3090 reaches 28.5 fps.After lightweighting, the parameter count decreased to 92M and the speed increased to 45.8 fps.This study effectively alleviates the cross modal semantic gap and feature redundancy issues, providing an efficient and accurate solution for intelligent video analysis.

In recent years, the task of translating from one language to another has attracted wide attention from researchers due to numerous practical uses, ranging from the translation of various texts and speeches, including the so-called "machine" translation, to the dubbing of films and numerous other video materials. To study this problem, we propose to use the information-theoretic method for assessing the quality of translations. We based our approach on the classification of sources of text variability proposed by A.N. Kolmogorov: information content, form, and unconscious author's style. It is clear that the unconscious "author's" style is influenced by the translator. So researchers need special methods to determine how accurately the author's style is conveyed, because it, in a sense, determines the quality of the translation. In this paper, we propose a method that allows us to estimate the quality of translation from different translators. The method is used to study translations of classical English-language works into Russian and, conversely, Russian classics into English. We successfully used this method to determine the attribution of literary texts.

Having recognized gender bias as a major issue affecting current translation technologies, researchers have primarily attempted to mitigate it by working on the data front. However, whether algorithmic aspects concur to exacerbate unwanted outputs remains so far under-investigated. In this work, we bring the analysis on gender bias in automatic translation onto a seemingly neutral yet critical component: word segmentation. Can segmenting methods influence the ability to translate gender? Do certain segmentation approaches penalize the representation of feminine linguistic markings? We address these questions by comparing 5 existing segmentation strategies on the target side of speech translation systems. Our results on two language pairs (English-Italian/French) show that state-of-the-art sub-word splitting (BPE) comes at the cost of higher gender bias. In light of this finding, we propose a combined approach that preserves BPE overall translation quality, while leveraging the higher ability of character-based segmentation to properly translate gender.

Multi-model classification method in heterogeneous image databases

In the coupling state of rotor unbalance fault and bearing defect fault for rotor system, the signals contain multiple fault components, and the fault diagnosis of the rotor system needs to contain comprehensive multidimensional feature quantities. However, irrelevant feature information in the multi-dimensional feature quantities increases the complexity of classification calculation and affects the efficiency and accuracy of diagnosis. In order to eliminate redundant and irrelevant features in the feature information, and achieve the goal of fewer diagnostic features and good diagnostic results, this paper proposes an adaptive feature selection based on the maximum information coefficient FF-FC-MIC (Feature-to-Feature and Feature-to-Category Maximum Information Coefficient) method. Firstly, the sparse representation algorithm is used to reconstruct the original signal to improve the signal-to-noise ratio, and the multi-dimensional feature quantity of the reconstructed signal is calculated; Secondly, calculate the correlation between features and features through MIC to obtain a feature set of weak correlation between features; thirdly, use MIC to calculate the correlation between features and signal categories to obtain a feature set with strong correlation between features and signals; Finally, the FF-FC-MIC feature selection method is used for feature adaptive selection and input into SVM to complete fault diagnosis. The method is analyzed by simulation signals and the real experiment signals. The results show that the method can effectively remove redundant and disclosed features in the coupling fault, reducing the characteristic dimension to reduce the fault classification time, and improve classification accuracy. Different experimental cases and various feature selection comparison methods further verify the accuracy and applicability of the proposed method.

Performing image retrieval in large image database is becoming popular with the development of multimedia technology. Recently, ISO/IEC moving pictures experts group (MPEG) drafts compact descriptors for visual search (CDVS) to support the related applications. The state-of-the-art feature selection strategy in CDVS adopts a priori relevance measure to guide the selection. However, this method ignores the gradient information of the described feature. In this paper, we adopt CDVS to address traffic vehicle search in large database. After detailedly analyzing to the local descriptor, we firstly define local descriptor distinctive degree based on the gradient quantity and the spatial gradient distribution energy of the feature. Then we propose our adaptive feature selection method by combing feature distinctive degree and the priori information. The proposed method is proven to be better than the standard algorithm in CDVS and it almost introduces 10% mean average precision (MAP) and top match rate increase in low bit rate mode.

Analyzing user behavior characteristics in a complex power grid environment is essential for user behavior planning and resource coordination optimization. Traditional user behavior analysis methods based on model-driven and causal analysis have the disadvantages of strong subjectivity and physical models that are difficult to deal with the randomness and uncertainty of user behavior in complex grid environments. In this paper, we use unsupervised learning methods to analyze user behavior in complex power grid environments, and propose user behavior analysis methods based on stacked autoencoder and clustering. We first reduce the complexity of user behavior data by proposing adaptive feature selection algorithm of user behavior based on stacked autoencoder and unsupervised learning (AFS-SAEUL). Finally, we build a user behavior analysis model based on adaptive feature selection and improved clustering (UBA-AFSIC). The model improved the performance of unsupervised classification of user behavior by fusing the adaptive generation strategy of the initial cluster centers. The simulation experiment results on two real electricity datasets and one public electric vehicle charging dataset show that compared with the existing feature selection algorithm and clustering algorithm, the algorithms proposed in this paper have higher feature selection rate and better clustering performance.

Recent studies on end-to-end speech translation(ST) have facilitated the exploration of multilingual end-to-end ST and end-to-end simultaneous ST. In this paper, we investigate end-to-end simultaneous speech translation in a one-to-many multilingual setting which is closer to applications in real scenarios. We explore a separate decoder architecture and a unified architecture for joint synchronous training in this scenario. To further explore knowledge transfer across languages, we propose an asynchronous training strategy on the proposed unified decoder architecture. A multi-way aligned multilingual end-to-end ST dataset was curated as a benchmark testbed to evaluate our methods. Experimental results demonstrate the effectiveness of our models on the collected dataset. Our codes and data are available at: https://github.com/XiaoMi/TED-MMST.

We proposed a new adaptive feature extraction (FEA) approach that integrates concepts of per-pixel/field classification and spectral ummixing. It combines their advantages in adaptive feature selection while minimizing the disadvantages associated with the high-complexity of each technique. The approach consists of local gradients calculation, reference clusters determination, prototype classification using fuzzy classifier, and feature vectors selection. Multiple experiments were performed using a simulated hyperspectral cube composed by 123 samples and 1254 features and classification was done only for verification purposes. Cross-validation demonstrated that FEA generated an average improvement of 7% on the misclassification error when compared to full feature analysis.

We present an effective method to progressively integrate and refine the cross-modality complementarities for RGB-D salient object detection (SOD). The proposed network mainly solves two challenging issues: 1) how to effectively integrate the complementary information from RGB image and its corresponding depth map, and 2) how to adaptively select more saliency-related features. First, we propose a cross-modality feature modulation (cmFM) module to enhance feature representations by taking the depth features as prior, which models the complementary relations of RGB-D data. Second, we propose an adaptive feature selection (AFS) module to select saliency-related features and suppress the inferior ones. The AFS module exploits multi-modality spatial feature fusion with the self-modality and cross-modality interdependencies of channel features are considered. Third, we employ a saliency-guided position-edge attention (sg-PEA) module to encourage our network to focus more on saliency-related regions. The above modules as a whole, called cmMS block, facilitates the refinement of saliency features in a coarse-to-fine fashion. Coupled with a bottom-up inference, the refined saliency features enable accurate and edge-preserving SOD. Extensive experiments demonstrate that our network outperforms state-of-the-art saliency detectors on six popular RGB-D SOD benchmarks.