Cross-subject EMG hand gesture recognition based on dynamic domain generalization.

Electromyography (EMG) signal based cross-subject gesture recognition methods reduce the influence of individual differences using transfer learning technology. These methods generally require calibration data collected from new subjects to adapt the pre-trained model to existing subjects. However, collecting calibration data is usually trivial and inconvenient for new subjects. This is currently a major obstacle to the daily use of hand gesture recognition based on EMG signals. To tackle the problem, we propose a novel dynamic domain generalization (DDG) method which is able to achieve accurate recognition on the hand gesture of new subjects without any calibration data. In order to extract more robust and adaptable features, a meta-adjuster is leveraged to generate a series of template coefficients to dynamically adjust dynamic network parameters. Specifically, two different kinds of templates are designed, in which the first one is different kinds of features, such as temporal features, spatial features, and spatial-temporal features, and the second one is different normalization layers. Meanwhile, a mix-style data augmentation method is introduced to make the meta-adjuster's training data more diversified. Experimental results on a public dataset verify that the proposed DDG outperforms the counterpart methods.

Vibroarthrography-based Knee Lesions Location via Multi-Label Embedding Learning.

Vibration arthrography (VAG) signals are widely utilized for knee pathology recognition due to their non-invasive and radiation-free nature. While most studies focus on determining knee health status, few have examined using VAG signals to locate knee lesions, which would greatly aid physicians in diagnosis and patient monitoring. To address this, we propose using Multi-Label classification (MLC) to efficiently locate different types of lesions within a single input. However, current MLC methods are not suitable for knee lesion location due to two major issues: 1) the positive-negative imbalance of pathological labels in knee pathology recognition is not considered, leading to poor performance, and 2) sparse label correlations between different lesions cannot be effectively extracted. Our solution is a label autoencoder incorporating a pre-trained model (PTM-LAE). To mitigate the positive-negative disequilibrium, we propose a pre-trained feature mapping model utilizing focal loss to dynamically adjust sample weights and focus on difficult-to-classify samples. To better explore the correlations between sparse labels, we introduce a Factorization-Machine-based neural network (DeepFM) that combines higher-order and lower-order correlations between different lesions. Experiments on our collected VAG data demonstrate that our model outperforms state-of-the-art methods.

Online Unsupervised Domain Adaptation via Reducing Inter- and Intra-Domain Discrepancies.

Unsupervised domain adaptation (UDA) transfers knowledge from a labeled source domain to an unlabeled target domain on cross-domain object recognition by reducing a distribution discrepancy between the source and target domains (interdomain discrepancy). Prevailing methods on UDA were presented based on the premise that target data are collected in advance. However, in online scenarios, the target data often arrive in a streamed manner, such as visual image recognition in daily monitoring, which means that there is a distribution discrepancy between incoming target data and collected target data (intradomain discrepancy). Consequently, most existing methods need to re-adapt the incoming data and retrain a new model on online data. This paradigm is difficult to meet the real-time requirements of online tasks. In this study, we propose an online UDA framework via jointly reducing interdomain and intradomain discrepancies on cross-domain object recognition where target data arrive in a streamed manner. Specifically, the proposed framework comprises two phases: classifier training and online recognition phases. In the former, we propose training a classifier on a shared subspace where there is a lower interdomain discrepancy between the two domains. In the latter, a low-rank subspace alignment method is introduced to adapt incoming data to the shared subspace by reducing the intradomain discrepancy. Finally, online recognition results can be obtained by the trained classifier. Extensive experiments on DA benchmarks and real-world datasets are employed to evaluate the performance of the proposed framework in online scenarios. The experimental results show the superiority of the proposed framework in online recognition tasks.

Reducing bias to source samples for unsupervised domain adaptation.

Unsupervised Domain Adaptation (UDA) makes predictions for the target domain data while labels are only available in the source domain. Lots of works in UDA focus on finding a common representation of the two domains via domain alignment, assuming that a classifier trained in the source domain can be generalized well to the target domain. Thus, most existing UDA methods only consider minimizing the domain discrepancy without enforcing any constraint on the classifier. However, due to the uniqueness of each domain, it is difficult to achieve a perfect common representation, especially when there is low similarity between the source domain and the target domain. As a consequence, the classifier is biased to the source domain features and makes incorrect predictions on the target domain. To address this issue, we propose a novel approach named reducing bias to source samples for unsupervised domain adaptation (RBDA) by jointly matching the distribution of the two domains and reducing the classifier's bias to source samples. Specifically, RBDA first conditions the adversarial networks with the cross-covariance of learned features and classifier predictions to match the distribution of two domains. Then to reduce the classifier's bias to source samples, RBDA is designed with three effective mechanisms: a mean teacher model to guide the training of the original model, a regularization term to regularize the model and an improved cross-entropy loss for better supervised information learning. Comprehensive experiments on several open benchmarks demonstrate that RBDA achieves state-of-the-art results, which show its effectiveness for unsupervised domain adaptation scenarios.

Online Cross-subject Emotion Recognition from ECG via Unsupervised Domain Adaptation.

Performing cross-subject emotion recognition (ER) using electrocardiogram (ECG) is challenging, since inter-subject discrepancy (caused by individual differences) between source and target subjects (new subjects) may hinder the generalization for new subjects. Recently, some ER methods based on unsupervised domain adaptation (UDA) are proposed to address inter-subject discrepancy. However, when being applied for online scenarios with time-varying ECG, existing methods may suffer performance degradation due to neglecting intra-subject discrepancy (caused by time-varying ECG) within target subjects, or need to re-train ER model, leading to time-and resource-consuming. In the paper, we propose an online cross-subject ER approach from ECG signals via UDA, consisting of two stages. In a training stage, we propose to train a classifier on a shared subspace with a lower inter-subject discrepancy. In an online recognition stage, an online data adaptation (ODA) method is introduced to adapt time-varying ECG via reducing the intra-subject discrepancy, and then online recognition results can be obtained by the trained classifier. Experimental results on Dreamer and Amigos with emotions of valence and arousal demonstrate that our proposed approach improves the classification accuracy by about 12% compared with the baseline method, and is robust to time-varying ECG in online scenarios.

End-to-End Versatile Human Activity Recognition with Activity Image Transfer Learning.

Transfer learning is a common solution to address cross-domain identification problems in Human Activity Recognition (HAR). Most existing approaches typically perform cross-subject transferring while ignoring transfers between different sensors or body parts, which limits the application scope of these models. Only a few approaches have been made to design a versatile HAR approach (cross-subject, cross-sensor and cross-body-part). Unfortunately, these existing approaches depend on complex handcrafted features and ignore the inequality of samples for positive transfer, which will hinder the transfer performance. In this paper, we propose a framework for versa-tile cross-domain activity recognition. Specifically, the proposed framework allows end-to-end implementation by exploiting adaptive features from activity image instead of extracting handcrafted features. And the framework uses a two-stage adaptation strategy consisting of pretraining stage and re-weighting stage to perform knowledge transfer. The pretraining stage ensures transferability of the source domain as well as separability of the target domain, and the re-weighting stage rebalances the contribution of the two domain samples. These two stages enhance the ability of knowledge transfer. We evaluate the performance of the proposed framework by conducting comprehensive experiments on three public HAR datasets (DSADS, OPPORTUNITY, and PAMAP2), and the experimental results demonstrate the effectiveness of our framework in versatile cross-domain HAR.

Cross-subject EEG-based Emotion Recognition Using Adversarial Domain Adaption with Attention Mechanism.

Cross-subject EEG-based emotion recognition (ER) is a rewarding work in real-life applications, due to individual differences between one subject and another subject. Most existing studies focus on training a subject-specific ER model. However, it is time-consuming and unrealistic to design the customized subject-specific model for a new subject in cross-subject scenarios. In this paper, we propose an Adversarial Domain Adaption with an Attention Mechanism method for EEG-based ER, namely ADAAM-ER, to decrease the individual discrepancy. ADAAM-ER consists of a Graph Convolution Neural Networks with CNNs (GCNN-CNNs) and an Adversarial Domain Adaption with a Level-wise Attention Mechanism (ADALAM). Specifically, GCNN-CNNs as a feature extractor, which constructs a broader feature space, is designed to obtain more discriminative features. And ADALAM, which can decrease the individual discrepancy by alignment of the more transferable feature regions, is introduced to further obtain the discriminative features with higher transferability. Consequently, the proposed ADAAM-ER method can design a more transferable emotion recognition model with more discriminative features for a new subject via improving transferability. Experimental results on the SEED dataset have verified the effectiveness of the proposed ADAAM-ER method with the mean accuracy of 86.58%.

PPG-based Biometric Identification: Discovering and Identifying a New User.

The convenience of Photoplethysmography (PPG) signal acquisition from wearable devices makes it becomes a hot topic in biometric identification. A majority of studies focus on PPG biometric technology in a verification application rather than an identification application. Yet, in the identification application, it is an inevitable problem in discovering and identifying a new user. However, so far few works have investigated this problem. Existing approaches can only identify trained old users. Their identification model needs to be retrained when a new user joins, which reduces the identification accuracy. This work investigates the approach and performance of identifying both old users and new users on a deep neural network trained only by old users. We used a deep neural network as a feature extractor, and the distance of the feature vector to discover and identify a new user, which avoids retraining the identification model. On the BIDMC data set, we achieved an accuracy of more than 99% for old users, an accuracy of more than 90% for discovering a new user, and an average accuracy of about 90% for identifying a new user. Our proposed approach can accurately identify old users and has feasibility in discovering and identifying a new user without retraining in the identification application.

Alleviating Feature Confusion in Cross-Subject Human Activity Recognition via Adversarial Domain Adaptation Strategy.

Sensor-based Human Activity Recognition (HAR) plays an important role in health care. However, great individual differences limit its application scenarios and affect its performance. Although general domain adaptation methods can alleviate individual differences to a certain extent, the performance of these methods is still not satisfactory, since the feature confusion caused by individual differences tends to be underestimated. In this paper, for the first time, we analyze the feature confusion problem in cross-subject HAR and summarize it into two aspects: Confusion at Decision Boundaries (CDB) and Confusion at Overlapping (COL). The CDB represents the misclassification caused by the feature located near the decision boundary, while the COL represents the misclassification caused by the feature aliasing of different classes. In order to alleviate CDB and COL to improve the stability of trained model when processing the data from new subjects, we propose a novel Adversarial Cross-Subject (ACS) method. Specifically, we design a parallel network that can extract features from both image space and time series simultaneously. Then we train two classifiers adversarially, and consider both features and decision boundaries to optimize the distribution to alleviate CDB. In addition, we introduce Minimum Class Confusion loss to reduce the confusion between classes to alleviate COL. The experiment results on USC-HAD dataset show that our method outperforms other generally used cross-subject methods.

Multi-label Arrhythmia Classification from Fixed-length Compressed ECG Segments in Real-time Wearable ECG Monitoring.

Recently, classification from compressed physiological signals in compressed sensing has been successfully applied to cardiovascular disease monitoring. However, in real-time wearable electrocardiogram (ECG) monitoring, it is very difficult to directly obtain the heartbeats information from compressed ECG signals. Thus arrhythmia classification from compressed ECG signals has to be handled in fixed-length segments instead of individual heartbeats. An inevitable issue is that a fixed-length ECG segment may contain multiple different types of arrhythmia. As a result, it is not appropriate to represent the multi-type real arrhythmia with a single label. In this paper, we first introduce multiple labels into fixed-length compressed ECG segments to challenge the arrhythmia classification issue. Then, we propose a deep learning model, which can directly classify multiple different types of arrhythmia from fixed-length compressed ECG segments with the advantages of low time cost for data processing and relatively high classification accuracy at a high compression ratio. Experimental results on the MIT-BIH arrhythmia database show that the exact match rate of our proposed method has reached 96.03% at CR(Compression Ratio)=70%, 94.99% at CR=80% and 93.19% at CR=90%.