Search

Graph neural networks (GNNs) have been highly successful in graph representation learning. The goal of GNNs is to enrich node representations by aggregating information from neighboring nodes. Much work has attempted to improve the quality of aggregation by introducing a variety of graph information with representational capabilities. The class of GNNs that improves the quality of aggregation by encoding graph information with representational capabilities into the weights of neighboring nodes through different learnable transformation structures (LTSs) are referred to as implicit GNNs. However, we argue that LTSs only transform graph information into the weights of neighboring nodes in the direction that minimizes the loss function during the learning process and does not actually utilize the effective properties of graph information, a phenomenon that we refer to as graph information vanishing (GIV). To validate this point, we perform thousands of experiments on seven node classification benchmark datasets. We first replace the graph information utilized by five implicit GNNs with random values and surprisingly observe that the variation range of accuracies is less than ± 0.3%. Then, we quantitatively characterize the similarity of the weights generated from graph information and random values by cosine similarity, and the cosine similarities are greater than 0.99. The empirical experiments show that graph information is equivalent to initializing the input of LTSs. We believe that graph information as an additional supervised signal to constrain the training of GNNs can effectively solve GIV. Here, we propose GinfoNN, which utilizes both labels and discrete graph curvature as supervised signals to jointly constrain the training of the model. The experimental results show that the classification accuracies of GinfoNN improve by two percentage points over baselines on large and dense datasets. [ABSTRACT FROM AUTHOR]

Copyright of Retos: Nuevas Perspectivas de Educación Física, Deporte y Recreación is the property of Federacion Espanola de Asociaciones de Docentes de Educacion Fisica 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.)

Accurately predicting the capacity and remaining useful life (RUL) of lithium-ion batteries is crucial for their reliable and safe functioning. In this research, we propose a dual Siamese transformer-encoder-based network consisting of two subnetworks to improve the RUL prediction capability and for wider applications. The first subnetwork, autoTrans, adopts a transformer-encoder architecture to form an autoencoder structure for the feature extraction. The second subnetwork, regTrans, is a transformer-encoder-based regressor model which takes the featuring encodings from autoTrans as inputs and makes the RUL prediction. To enhance the model's robustness, both subnetworks employ the Siamese architecture for handling raw and noisy inputs. A joint training strategy is applied on autoTrans and regTrans to optimize the proposed approach. The experimental verification is conducted on the NASA battery, and our model achieves the best average results across different evaluation criteria. Furthermore, we also apply our model to the turbofan engine dataset and demonstrate promising performance as well. [ABSTRACT FROM AUTHOR]

Distant supervised relation extraction has been increasingly popular in recent years, which generates datasets automatically without human intervention. However, the distant supervised assumption has the limitation that the generated datasets have inevitable labeling errors. This paper proposes the method of Hierarchical Symmetric Cross Entropy for Distant Supervised Relation Extraction (HSCERE) to alleviate the impact of the noisy labels. Specifically, HSCERE simultaneously utilizes two extractors with the same network structure for collaborative learning. This collaborative learning process guides the optimization of the extractor through a joint loss function, namely Hierarchical Symmetric Cross Entropy (HSCE). Within the HSCE loss, the predicted probability distribution of the extractors serves as the supervisory signal, guiding the optimization of the extractors on two levels to reduce the impact of noisy labels. The two levels include the internal optimization within each extractor and the collaborative optimization between extractors. Experiments on generally used datasets show that HSCERE can effectively handle noisy labels and can be incorporated into various methods to enhance their performance. [ABSTRACT FROM AUTHOR]

The intending goal of modeling and simulation is to create a unified Live-Virtual-Constructive (LVC) integration architecture, support rapid integration model and carry out simulation for joint flight training, tactical coordination, operational planning and evaluation. By exploring the connotation and application field of LVC, this paper analyzes the development plan of LVC training network, combs three typical application examples of LVC system, which are distributed mission combat, joint simulation environment and US Navy simulation training. Using the experience of LVC to solve the joint training, this paper gives the enlightenment of accelerating the construction of the joint training of our army, and provides valuable reference for the construction of our army's joint training environment.

Noisy labels due to mistakes in manual labeling or data collecting are challenging for the expansion of deep neural network applications. Current robust network learning methods such as Decoupling, Co-teaching, and Joint Training with Co-Regularization are very promising for learning with noisy labels, yet the coordination between consistency and diversity is not fully considered which is crucial for the performance of the model. To tackle this issue, a novel robust learning paradigm called Joint training by combining Consistency and Diversity (JoCaD) is proposed in this paper. The JoCaD is devoted to maximize the prediction consistency of the networks while keeping enough diversity on their representation learning. Specifically, aiming to reconcile the relationship between consistency and diversity, an effective implementation is proposed which dynamically adjusts joint loss to boost the model learning with noisy labels. The extensive experimental results on MNIST, CIFAR-10, CIFAR-100, and Clothing1M demonstrate that our proposed JoCaD has better performance than some representative SOTA methods. [ABSTRACT FROM AUTHOR]

One potential way to enhance the performance of facial expression recognition (FER) is to augment the training set by increasing the number of samples. By incorporating multiple FER datasets, deep learning models can extract more discriminative features. However, the inconsistent labeling criteria and subjective biases found in annotated FER datasets can significantly hinder the recognition accuracy of deep learning models when handling mixed datasets. Effectively perform joint training on multiple datasets remains a challenging task. In this study, we propose a joint training method for training an FER model using multiple FER datasets. Our method consists of four steps: (1) selecting a subset from the additional dataset, (2) generating pseudo-continuous labels for the target dataset, (3) refining the labels of different datasets using continuous label mapping and discrete label relabeling according to the labeling criteria of the target dataset, and (4) jointly training the model using multi-task learning. We conduct joint training experiments on two popular in-the-wild FER benchmark databases, RAF-DB and CAER-S, while utilizing the AffectNet dataset as an additional dataset. The experimental results demonstrate that our proposed method outperforms the direct merging of different FER datasets into a single training set and achieves state-of-the-art performance on RAF-DB and CAER-S with accuracies of 92.24% and 94.57%, respectively. [ABSTRACT FROM AUTHOR]

The intending goal of modeling and simulation is to create a unified Live-Virtual-Constructive (LVC) integration architecture, support rapid integration model and carry out simulation for joint flight training, tactical coordination, operational planning and evaluation. By exploring the connotation and application field of LVC, this paper analyzes the development plan of LVC training network, combs three typical application examples of LVC system, which are distributed mission combat, joint simulation environment and US Navy simulation training. Using the experience of LVC to solve the joint training, this paper gives the enlightenment of accelerating the construction of the joint training of our army, and provides valuable reference for the construction of our army's joint training environment. [ABSTRACT FROM AUTHOR]

Existing unsupervised person re-identification approaches fail to fully capture the fine-grained features of local regions, which can result in people with similar appearances and different identities being assigned the same label after clustering. The identity-independent information contained in different local regions leads to different levels of local noise. To address these challenges, joint training with local soft attention and dual cross-neighbor label smoothing (DCLS) is proposed in this study. First, the joint training is divided into global and local parts, whereby a soft attention mechanism is proposed for the local branch to accurately capture the subtle differences in local regions, which improves the ability of the re-identification model in identifying a person's local significant features. Second, DCLS is designed to progressively mitigate label noise in different local regions. The DCLS uses global and local similarity metrics to semantically align the global and local regions of the person and further determines the proximity association between local regions through the cross information of neighboring regions, thereby achieving label smoothing of the global and local regions throughout the training process. In extensive experiments, the proposed method outperformed existing methods under unsupervised settings on several standard person re-identification datasets. [ABSTRACT FROM AUTHOR]

The integration of graph embedding technology and collaborative filtering algorithms has shown promise in enhancing the performance of recommendation systems. However, existing integrated recommendation algorithms often suffer from feature bias and lack effectiveness in personalized user recommendation. For instance, users' historical interactions with a certain class of items may inaccurately lead to recommendations of all items within that class, resulting in feature bias. Moreover, accommodating changes in user interests over time poses a significant challenge. This study introduces a novel recommendation model, RCKFM, which addresses these shortcomings by leveraging the CoFM model, TransR graph embedding model, backdoor tuning of causal inference, KL divergence, and the factorization machine model. RCKFM focuses on improving graph embedding technology, adjusting feature bias in embedding models, and achieving personalized recommendations. Specifically, it employs the TransR graph embedding model to handle various relationship types effectively, mitigates feature bias using causal inference techniques, and predicts changes in user interests through KL divergence, thereby enhancing the accuracy of personalized recommendations. Experimental evaluations conducted on publicly available datasets, including "MovieLens-1M" and "Douban dataset" from Kaggle, demonstrate the superior performance of the RCKFM model. The results indicate a significant improvement of between 3.17% and 6.81% in key indicators such as precision, recall, normalized discount cumulative gain, and hit rate in the top-10 recommendation tasks. These findings underscore the efficacy and potential impact of the proposed RCKFM model in advancing recommendation systems. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Xinjiang University (Natural Science Edition) is the property of Xinjiang University 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.)

This paper addresses a joint training approach applied to a pipeline comprising speech enhancement (SE) and automatic speech recognition (ASR) models, where an acoustic tokenizer is included in the pipeline to leverage the linguistic information from the ASR model to the SE model. The acoustic tokenizer takes the outputs of the ASR encoder and provides a pseudo-label through K-means clustering. To transfer the linguistic information, represented by pseudo-labels, from the acoustic tokenizer to the SE model, a cluster-based pairwise contrastive (CBPC) loss function is proposed, which is a self-supervised contrastive loss function, and combined with an information noise contrastive estimation (infoNCE) loss function. This combined loss function prevents the SE model from overfitting to outlier samples and represents the pronunciation variability in samples with the same pseudo-label. The effectiveness of the proposed CBPC loss function is evaluated on a noisy LibriSpeech dataset by measuring both the speech quality scores and the word error rate (WER). The experimental results reveal that the proposed joint training approach using the described CBPC loss function achieves a lower WER than the conventional joint training approaches. In addition, it is demonstrated that the speech quality scores of the SE model trained using the proposed training approach are higher than those of the standalone-SE model and SE models trained using conventional joint training approaches. An ablation study is also conducted to investigate the effects of different combinations of loss functions on the speech quality scores and WER. Here, it is revealed that the proposed CBPC loss function combined with infoNCE contributes to a reduced WER and an increase in most of the speech quality scores. [ABSTRACT FROM AUTHOR]

Speech data gathered from real-world environments typically contain noise, a significant element that undermines the performance of deep neural network-based speaker verification (SV) systems. To mitigate performance degradation due to noise and develop noise-robust SV systems, several researchers have integrated speech enhancement (SE) and SV systems. We previously proposed the extended U-Net (ExU-Net), which achieved state-of-the-art performance in SV in noisy environments by jointly training SE and SV systems. In the SE field, some studies have shown that recognizing noise components within speech can improve the system’s performance. Inspired by these approaches, we propose a noise-aware ExU-Net (NA-ExU-Net) that acknowledges noise information in the SE process based on the ExU-Net architecture. The proposed system comprises a Split Encoder and a feature refinement module (FRM). The Split Encoder handles the speech and noise separately by dividing the encoder blocks, whereas FRM is designed to inhibit the propagation of irrelevant data via skip connections. To validate the effectiveness of our proposed framework in noisy conditions, we evaluated the models on the VoxCeleb1 test set with added noise from the MUSAN corpus. The experimental results demonstrate that NA-ExU-Net outperforms the ExU-Net and other baseline systems under all evaluation conditions. Furthermore, evaluations in out-of-domain noise environments indicate that NA-ExU-Net significantly surpasses existing frameworks, highlighting its robustness and generalization capabilities. The codes utilized in our experiments can be accessed at https://github.com/chan-yeong0519/NA-ExU-Net.

As the problem of information overload becomes more severe, it has become increasingly difficult for users to browse news that they are interested in. News recommendation is an effective method to alleviate this problem. In news recommendation, accurately obtaining rich semantic news representations and modeling user’s historical interests are long-standing challenges. Based on these two issues, we propose a News recommendation approach with Word-related joint Topic prediction(NWT). The core of our approach is the topic perceptron and the news encoder that emphasizes word relevance. The topic information and word relevance information are fused through joint training, enhancing the abilities of topic prediction and news representation. In the news encoder, we employ multi-head self-attention network to emphasize word relevance, capturing the semantic relationships between any two words in the news title and enhancing the representation of the news. In the topic perceptron, we use the enhanced word relevance-based news representation as input and learn the topic information in the news by assigning weights to different topics. During joint training, the overall news recommendation module obtains the news topic information. Additionally, the user encoder learns the user’s topic preferences and utilizes attention networks to highlight news articles that are more representative of the user’s interests, thereby acquiring a more accurate user representation. We conduct extensive experiments on the MIND dataset, and the results demonstrate that NWT outperforms most existing baseline methods across various evaluation metrics.

Abstract Target speaker separation aims to separate the speech components of the target speaker from mixed speech and remove extraneous components such as noise. In recent years, deep learning-based speech separation methods have made significant breakthroughs and have gradually become mainstream. However, these existing methods generally face problems with system latency and performance upper limits due to the large model size. To solve these problems, this paper proposes improvements in the network structure and training methods to enhance the model’s performance. A lightweight target speaker separation network based on long-short-term memory (LSTM) is proposed, which can reduce the model size and computational delay while maintaining the separation performance. Based on this, a target speaker separation method based on joint training is proposed to achieve the overall training and optimization of the target speaker separation system. Joint loss functions based on speaker registration and speaker separation are proposed for joint training of the network to further improve the system’s performance. The experimental results show that the lightweight target speaker separation network proposed in this paper has better performance while being lightweight, and joint training of the target speaker separation network with our proposed loss function can further improve the separation performance of the original model.

Graph neural networks (GNNs) have been highly successful in graph representation learning. The goal of GNNs is to enrich node representations by aggregating information from neighboring nodes. Much work has attempted to improve the quality of aggregation by introducing a variety of graph information with representational capabilities. The class of GNNs that improves the quality of aggregation by encoding graph information with representational capabilities into the weights of neighboring nodes through different learnable transformation structures (LTSs) are referred to as implicit GNNs. However, we argue that LTSs only transform graph information into the weights of neighboring nodes in the direction that minimizes the loss function during the learning process and does not actually utilize the effective properties of graph information, a phenomenon that we refer to as graph information vanishing (GIV). To validate this point, we perform thousands of experiments on seven node classification benchmark datasets. We first replace the graph information utilized by five implicit GNNs with random values and surprisingly observe that the variation range of accuracies is less than ± 0.3%. Then, we quantitatively characterize the similarity of the weights generated from graph information and random values by cosine similarity, and the cosine similarities are greater than 0.99. The empirical experiments show that graph information is equivalent to initializing the input of LTSs. We believe that graph information as an additional supervised signal to constrain the training of GNNs can effectively solve GIV. Here, we propose GinfoNN, which utilizes both labels and discrete graph curvature as supervised signals to jointly constrain the training of the model. The experimental results show that the classification accuracies of GinfoNN improve by two percentage points over baselines on large and dense datasets.

In existing sequential recommendation systems, user behavior data are directly used as training data for the model to complete the training process and address recommendation tasks. However, user-generated behavioral data inevitably contains noise, and the use of the Transformer's recommendation model may lead to overfitting on such noisy data. To address this issue, we introduce a sequence recommendation algorithm model named FAT-Rec, which incorporates fusion filters and converters through joint training. By employing joint training methods, we establish both a transformer prediction layer and a CTR prediction layer. Toward the end of the model, we assign weights and sum up the losses from the Transformer and CTR prediction layers to derive the final loss function. Experimental results on two widely used datasets, MovieLens and Goodbooks, demonstrate a significant enhancement in the performance of the proposed FAT-Rec recommendation algorithm compared with seven comparative models. This validates the efficacy of the fusion filter and transformer within the context of sequence recommendation tasks under the joint training mechanism. [ABSTRACT FROM AUTHOR]

Recently, advancements in machine-learning technology have enabled platforms such as short video applications and e-commerce websites to accurately predict user behavior and cater to their interests. However, the limited nature of user data may compromise the accuracy of these recommendation systems. To address personalized recommendation challenges and adapt to changes in user preferences, reinforcement-learning algorithms have been developed. These algorithms strike a balance between exploring new items and exploiting existing ones, thereby enhancing recommendation accuracy. Nevertheless, the cold-start problem and data sparsity continue to impede the development of these recommendation systems. Hence, we proposed a joint-training algorithm that combined deep reinforcement learning with dynamic user groups. The goal was to capture user preferences for precise recommendations while addressing the challenges of data sparsity and cold-start. We used embedding layers to capture representations and make decisions before the reinforcement-learning process, executing this approach cyclically. Through this method, we dynamically obtained more accurate user and item representations and provide precise recommendations. Additionally, to address data sparsity, we introduced a dynamic user grouping algorithm that collectively enhanced the recommendations using group parameters. We evaluated our model using movie-rating and e-commerce datasets. As compared to other baseline algorithms, our algorithm not only improved recommendation accuracy but also enhanced diversity by uncovering recommendations across more categories. [ABSTRACT FROM AUTHOR]

Automatic speech recognition (ASR) systems have emerged as indispensable tools across a wide spectrum of applications, ranging from transcription services to voice-activated assistants. To enhance the performance of these systems, it is important to deploy efficient models capable of adapting to diverse deployment conditions. In recent years, on-demand pruning methods have obtained significant attention within the ASR domain due to their adaptability in various deployment scenarios. However, these methods often confront substantial tradeoffs, particularly in terms of unstable accuracy when reducing the model size. To address challenges, this study introduces two crucial empirical findings. Firstly, it proposes the incorporation of an online distillationmechanism during on-demand pruning training, which holds the promise of maintaining more consistent accuracy levels. Secondly, it proposes the utilization of the Mogrifier long short-term memory (LSTM) language model (LM), an advanced iteration of the conventional LSTM LM, as an effective alternative for pruning targets within the ASR framework. Through rigorous experimentation on the ASR system, employing the Mogrifier LSTM LM and training it using the suggested joint on-demand pruning and online distillation method, this study provides compelling evidence. The results exhibit that the proposed methods significantly outperform a benchmark model trained solely with on-demand pruning methods. Impressively, the proposed strategic configuration successfully reduces the parameter count by approximately 39%, all the while minimizing trade-offs. [ABSTRACT FROM AUTHOR]

Target speaker separation aims to separate the speech components of the target speaker from mixed speech and remove extraneous components such as noise. In recent years, deep learning-based speech separation methods have made significant breakthroughs and have gradually become mainstream. However, these existing methods generally face problems with system latency and performance upper limits due to the large model size. To solve these problems, this paper proposes improvements in the network structure and training methods to enhance the model's performance. A lightweight target speaker separation network based on long-short-term memory (LSTM) is proposed, which can reduce the model size and computational delay while maintaining the separation performance. Based on this, a target speaker separation method based on joint training is proposed to achieve the overall training and optimization of the target speaker separation system. Joint loss functions based on speaker registration and speaker separation are proposed for joint training of the network to further improve the system's performance. The experimental results show that the lightweight target speaker separation network proposed in this paper has better performance while being lightweight, and joint training of the target speaker separation network with our proposed loss function can further improve the separation performance of the original model. [ABSTRACT FROM AUTHOR]

Semantic segmentation is important for the accuracy of target detection. Semantic labels are difficult to obtain for real driving; however, they are easy to obtain in virtual datasets. So this paper presents an adaptive joint training strategy based on real and virtual datasets: (1) building multi-modal fusion networks using image, depth and semantic information. (2) A joint training strategy of virtual and real datasets and data sharing is used for semantic information, and an adaptive optimizer is provided. The monocular detection network obtained by training with this strategy has a large improvement in its effectiveness relative to the conventional network. [ABSTRACT FROM AUTHOR]

The unsupervised domain-adaptive vehicle re-identification approach aims to transfer knowledge from a labeled source domain to an unlabeled target domain; however, there are knowledge differences between the target domain and the source domain. To mitigate domain discrepancies, existing unsupervised domain-adaptive re-identification methods typically require access to source domain data to assist in retraining the target domain model. However, for security reasons, such as data privacy, data exchange between different domains is often infeasible in many scenarios. To this end, this paper proposes an unsupervised domain-adaptive vehicle re-identification method based on source-free knowledge transfer. First, by constructing a source-free domain knowledge migration module, the target domain is consistent with the source domain model output to train a generator to generate the "source-like samples". Then, it can effectively reduce the model knowledge difference and improve the model's generalization performance. In the experiment, two mainstream public datasets in this field, VeRi776 and VehicleID, are tested experimentally, and the obtained rank-k (the cumulative matching features) and mAP (the mean Average Precision) indicators are both improved, which are suitable for object re-identification tasks when data between domains cannot be interoperated. [ABSTRACT FROM AUTHOR]

In open-domain question answering (QA), the system needs to answer questions from various fields and forms according to given passages. Machine reading comprehension (MRC) can assist the system in comprehending passages and questions, hence often used in improving the performance of the QA system. More passages will bring more features to the result of comprehension, making the reading result more accurate. However, there is still a lack of effective ways when facing multiple passages because it is difficult to use the effective information brought by multiple passages while processing noisy information. This study introduces a new MRC model called RRQA (Reconfirmed Reader for Open-Domain Question Answering) to integrate the QA scenario into traditional MRC, which focuses on the interaction between questions and passages, considering the noisy information. The proposed model consists of two parts: answer extraction and answer verification. In the part of answer extraction, a multi-layer neural network model extracts candidate answers from each passage according to the question, taking into account that the question may have no answer. In the answer verification part, the final answer is verified by combining the features of all candidate answers, which can reduce the incompleteness of the independent answer extraction. Experiments show that the model RRQA outperforms the state-of-the-art models on the WebQA dataset. [ABSTRACT FROM AUTHOR]

The integration of graph embedding technology and collaborative filtering algorithms has shown promise in enhancing the performance of recommendation systems. However, existing integrated recommendation algorithms often suffer from feature bias and lack effectiveness in personalized user recommendation. For instance, users’ historical interactions with a certain class of items may inaccurately lead to recommendations of all items within that class, resulting in feature bias. Moreover, accommodating changes in user interests over time poses a significant challenge. This study introduces a novel recommendation model, RCKFM, which addresses these shortcomings by leveraging the CoFM model, TransR graph embedding model, backdoor tuning of causal inference, KL divergence, and the factorization machine model. RCKFM focuses on improving graph embedding technology, adjusting feature bias in embedding models, and achieving personalized recommendations. Specifically, it employs the TransR graph embedding model to handle various relationship types effectively, mitigates feature bias using causal inference techniques, and predicts changes in user interests through KL divergence, thereby enhancing the accuracy of personalized recommendations. Experimental evaluations conducted on publicly available datasets, including “MovieLens-1M” and “Douban dataset” from Kaggle, demonstrate the superior performance of the RCKFM model. The results indicate a significant improvement of between 3.17% and 6.81% in key indicators such as precision, recall, normalized discount cumulative gain, and hit rate in the top-10 recommendation tasks. These findings underscore the efficacy and potential impact of the proposed RCKFM model in advancing recommendation systems.

This paper addresses a joint training approach applied to a pipeline comprising speech enhancement (SE) and automatic speech recognition (ASR) models, where an acoustic tokenizer is included in the pipeline to leverage the linguistic information from the ASR model to the SE model. The acoustic tokenizer takes the outputs of the ASR encoder and provides a pseudo-label through K-means clustering. To transfer the linguistic information, represented by pseudo-labels, from the acoustic tokenizer to the SE model, a cluster-based pairwise contrastive (CBPC) loss function is proposed, which is a self-supervised contrastive loss function, and combined with an information noise contrastive estimation (infoNCE) loss function. This combined loss function prevents the SE model from overfitting to outlier samples and represents the pronunciation variability in samples with the same pseudo-label. The effectiveness of the proposed CBPC loss function is evaluated on a noisy LibriSpeech dataset by measuring both the speech quality scores and the word error rate (WER). The experimental results reveal that the proposed joint training approach using the described CBPC loss function achieves a lower WER than the conventional joint training approaches. In addition, it is demonstrated that the speech quality scores of the SE model trained using the proposed training approach are higher than those of the standalone-SE model and SE models trained using conventional joint training approaches. An ablation study is also conducted to investigate the effects of different combinations of loss functions on the speech quality scores and WER. Here, it is revealed that the proposed CBPC loss function combined with infoNCE contributes to a reduced WER and an increase in most of the speech quality scores.

Copyright of Computer Measurement & Control is the property of Magazine Agency of Computer Measurement & Control 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.)

State-of-the-art performances for natural language processing tasks are achieved by supervised learning, specifically, by fine-tuning pre-trained language models such as BERT (Bidirectional Encoder Representation from Transformers). With increasingly accurate models, the size of the fine-tuned pre-training corpus is becoming larger and larger. However, very few studies have explored the selection of pre-training corpus. Therefore, this paper proposes a data enhancement-based domain pre-training method. At first, a pre-training task and a downstream fine-tuning task are jointly trained to alleviate the catastrophic forgetting problem generated by existing classical pre-training methods. Then, based on the hard-to-classify texts identified from downstream tasks' feedback, the pre-training corpus can be reconstructed by selecting the similar texts from it. The learning of the reconstructed pre-training corpus can deepen the model's understanding of undeterminable text expressions, thus enhancing the model's feature extraction ability for domain texts. Without any pre-processing of the pre-training corpus, the experiments are conducted for two tasks, named entity recognition (NER) and text classification (CLS). The results show that learning the domain corpus selected by the proposed method can supplement the model's understanding of domain-specific information and improve the performance of the basic pre-training model to achieve the best results compared with other benchmark methods. [ABSTRACT FROM AUTHOR]

Aspect and opinion pair extraction aims to extract aspect and opinion items and match relations from a given sentence. However, related studies typically extract aspects and opinions independently without identifying the relationships. To identify the relationships of aspect and opinion item, this paper proposed a knowledge-augmented multi-task learning model for aspect and opinion pair extraction. First, it used the pre-trained language model to generate word vectors with semantic information for the text. In order to achieve the effect of knowledge enhancement, it used the masked attention mechanism to integrate the semantic information of the knowledge graph into the word vectors, and used the sequence labeling method based on the distance-based attention and conditional random fields to extract aspects and opinions. Finally, it matched the extracted aspects and opinions to predict the corresponding relationship. In order to strengthen the connection between the aspect and opinion extraction module and the matching module, the model adopted a shared coding layer to achieve joint training. In addition, in the training process, the matching module used the real labels as input, and used the result of the extraction module as input in the testing process. Finally, to demonstrate the effectiveness of the model, this paper used three general domain datasets for comparative experiments. The model achieves F, values of 66. 99%, 75. 17% and 67. 30% in aspect and opinion matching tasks respectively, and outperforms other comparative models. [ABSTRACT FROM AUTHOR]

Inspired by the observation that pathologists pay more attention to the nuclei regions when analyzing pathological images, this study utilized soft segmentation to imitate the visual focus mechanism and proposed a new segmentation–classification joint model to achieve superior classification performance for breast cancer pathology images. Aiming at the characteristics of different sizes of nuclei in pathological images, this study developed a new segmentation network with excellent cross-scale description ability called DIU-Net. To enhance the generalization ability of the segmentation network, that is, to avoid the segmentation network from learning low-level features, we proposed the Complementary Color Conversion Scheme in the training phase. In addition, due to the disparity between the area of the nucleus and the background in the pathology image, there is an inherent data imbalance phenomenon, dice loss and focal loss were used to overcome this problem. In order to further strengthen the classification performance of the model, this study adopted a joint training scheme, so that the output of the classification network can not only be used to optimize the classification network itself, but also optimize the segmentation network. In addition, this model can also provide the pathologist model's attention area, increasing the model's interpretability. The classification performance verification of the proposed method was carried out with the BreaKHis dataset. Our method obtains binary/multi-class classification accuracy 97.24/93.75 and 98.19/94.43 for 200× and 400× images, outperforming existing methods. [ABSTRACT FROM AUTHOR]

Labeled data in insufficient amounts and missing categories are two observable features for high maneuvering target trajectory data. However, the existing research achievements are insufficient for solving these two problems simultaneously during data classification. This study proposed a two-stage semi-supervised trajectory data classification algorithm. By pre-training the autoencoder and combining it with the Siamese network, a two-stage joint training was formed, which enabled the model to deal with missing categories by clustering and maintaining the classification ability under the missing label categories. The experimental simulation results showed that the performance of this algorithm was better than the classical semi-supervised algorithm label propagation and transferred learning when the amount of various labeled data was as low as 1–5. The two-stage training model also had a good effect on the problem of missing categories. When 75% of the types were missing, the purity could still reach 82%, which was about eight percentage points higher than the directly trained network. When two problems appeared simultaneously, compared with the directly trained network, the performance improved by about three percentage points on average, and the purity was consistently higher than the clustering results. In summary, this algorithm was more tolerant of the problems of labeled datasets, so it was more practical. [ABSTRACT FROM AUTHOR]

In modern education, personalized learning has garnered increasing attention, with cognitive diagnosis (CD) serving as a fundamental component. The objective of CD is to deeply comprehend students' cognitive abilities, thereby identifying individual differences and providing support for personalized learning. However, many existing models ignore the influence of data noise and confidence in students' responses. In addition, most existing methods equally extract evidence of students' strengths and weaknesses from their overall responses, disregarding the characteristics of various types of relationships in students' correct and incorrect responses. To this end, this study proposes a Confidence-guided Multigraph Model (CMM) for CD. First, to capture the changes and uncertainties in the student's cognitive process, CMM dynamically calculates the confidence scores of the students' responses. Second, CMM constructs multiple graphs to represent various relationships in students' correct and incorrect responses. Then, CMM aggregates the unique information within the students' correct and incorrect responses from various graphs. Finally, CMM uses a joint training strategy of multigraph iteration to simulate students' cognitive ability in multiple dimensions. Extensive experimental results on two real-world datasets clearly demonstrate the efficacy of our proposed model. This innovative approach provides a novel perspective for the development of CD and lays the foundation for more accurate personalized learning support. • We considered confidence information to reduce the impact of data noise and sparsity. • A model using multi-graph to represent different types of relationships is developed. • We design a joint training strategy to integrate multiple levels of information. • Our model achieves more precise and robust results in diagnosis results. [ABSTRACT FROM AUTHOR]

The unsupervised domain-adaptive vehicle re-identification approach aims to transfer knowledge from a labeled source domain to an unlabeled target domain; however, there are knowledge differences between the target domain and the source domain. To mitigate domain discrepancies, existing unsupervised domain-adaptive re-identification methods typically require access to source domain data to assist in retraining the target domain model. However, for security reasons, such as data privacy, data exchange between different domains is often infeasible in many scenarios. To this end, this paper proposes an unsupervised domain-adaptive vehicle re-identification method based on source-free knowledge transfer. First, by constructing a source-free domain knowledge migration module, the target domain is consistent with the source domain model output to train a generator to generate the “source-like samples”. Then, it can effectively reduce the model knowledge difference and improve the model’s generalization performance. In the experiment, two mainstream public datasets in this field, VeRi776 and VehicleID, are tested experimentally, and the obtained rank-k (the cumulative matching features) and mAP (the mean Average Precision) indicators are both improved, which are suitable for object re-identification tasks when data between domains cannot be interoperated.

We propose a joint training scheme of an any-to-one voice conversion (VC) system with LPCNet to improve the speech naturalness, speaker similarity, and intelligibility of the converted speech. Recent advancements in neural-based vocoders, such as LPCNet, have enabled the production of more natural and clear speech. However, other components in typical VC systems are often designed independently, such as the conversion model. Hence, separate training strategies are used for each component that is not in direct correlation to the training objective of the vocoder preventing exploitation of the full potential of LPCNet. This problem is addressed by proposing a jointly trained conversion model and LPCNet. To accurately capture the linguistic contents of the given utterance, we use speaker-independent (SI) features derived from an automatic speech recognition (ASR) model trained using a mixed-language speech corpus. Subsequently, a conversion model maps the SI features to the acoustic representations used as input features to LPCNet. The possibility to synthesize cross-language speech using the proposed approach is also explored in this paper. Experimental results show that the proposed model can achieve real-time VC, unlocking the full potential of LPCNet and outperforming the state of the art.

To improve the precision of wind farm multi-machine equivalence and multi-scene generalization, this paper proposes a method for wind turbine clustering and equivalent parameter identification in multi-time scales based on the deep migration of multi-view features. The proposed technique carries out multi-machine equivalence by leveraging the multi-view information from each turbine in a wind farm. Specifically, a deep spatio-temporal Improved Auto-Encoder is designed, jointly trained with the target clustering layer. IAE is used for mining multi-view latent characteristics of wind turbines orienting to grouping turbines to improve the model’s adaptability to multiple scenarios and divide turbines in an unsupervised manner. This method generates a visual heat map to represent the attended area of characteristics based on transfer learning and Class Activation Map to enable interpretability. In the next phase, this technique constructs a multi-objective optimization model by synthesizing the equivalent deviation of voltage, current, active power, and reactive power to further improve accuracy. It can identify the equivalent parameters of collector lines, the mechanical structure, and the control system at different time scales simultaneously via the black-box paralleled optimization method based on Bayes and Multi-arm Bandit. The proposed approach is evaluated on a typical double-fed wind farm with grid-side faults under various conditions of disturbing winds. Also, an ablation study is conducted to make analysis according to the two phases, i.e., turbine division and parameter identification. The results validate the accuracy and robustness of this method.