Performing better text translation by integrating auxiliary inputs from visual information has gained widespread attention in recent years. While existing methods outperform the text-only translation models, the step-by-step generative style reduces the inference speed, which limits their applicability in real-world scenarios. In this paper, we propose the non-autoregressive language model (NA-LM) for multimodal machine translation. With NA-LM, we develop a Non-Autoregressive Multimodal Transformer (NA-MMT), which accelerates the generative translation via a parallel multimodal decoder. To retain the translation performance, we improve the NA-MMT in twofold: 1) We preprocess the image into a refined sequence of visual entities with length encoding to reduce irrelevant information; 2) We design cross fertility and cross-modal gate attention for multimodal decoder to enhance the generative quality. Experiments on Multi30k datasets show the NA-MMT can generate high-quality translation with over 11× speedup than the baselines, which is strongly competitive.

Unsupervised domain adaptation (UDA) methods have recently been explored for their use in skeleton recognition tasks. Much work along this line has been focusing on the “close-set” problems, which often deviate from reality as human actions vary in application scenarios. Thus, there remains a need to thoroughly study the “open-set” problems with UDA methods for skeleton recognition, aiming to support those models capable of self-adapting to action changes in different scenarios. To this end, we delve into the “open-set” problems from a feature alignment perspective under UDA settings in reaching domain and class alignment. Specifically, the domain-wise alignment was achieved by the maximum mean discrepancy (MMD) combined with supervision signals from the source domain, which form clear feature boundaries between the “known” and “unknown” classes. Then, the class-wise alignment was achieved by contrastive learning methods, which are distinguished from previous binary classification methods, in reaching compactness inside of “unknown” or “known” classes. Moreover, we conducted the Fourier analysis during the evaluation phases to verify the model’s robustness. To our knowledge, we are the first to apply the Fourier Heatmap in UDA methods for skeleton recognition. The heatmap visualizes the model’s sensitivity steered for interpretability. Significant performance improvements are observed on the NTU and PKU data sets when adding the domain-wise alignment module to other contrastive learning methods. Furthermore, experimental results demonstrate that our approach, termed CStrCRL-UDA, is consistent with robustness and efficiency on these two benchmark data sets.

Most existing video-based posture recognition methods treat frames equally using unified or random sampling strategies, thus losing the temporal relationship information among frames. To address this problem, we propose a lightweight framework, namely RASNet, to adaptively select informative frames for recognition. Specifically, we design a video-suited exploration environment to guide the agent in learning the selection strategy. We introduce the reparametrization method to convert the discrete action space into a continuous space, making the agent robust and random. For the reward part, we design a multi-factor function to reward the agent keeping a balance between frame usage and accuracy. Extensive experiments on three large-scale datasets prove the effectiveness of RASNet, e.g., achieving 85.9% accuracy with fewer 1.15 frames than other state-of-the-art methods on Kinetics 600.

Recently, the intelligent question answering (QA) system for home appliances has attracted wide attention in China because it can provide users with reasonable suggestions in time.However, the fuzzy boundaries of the Chinese "word groups" and the absence of a unified standard for Chinese question classification result in low efficiency of QA, which has hindered the development of the system for a long time.In this work, we developed a QA model based on RoBERTa-WWM to answer questions about home appliances in specific fields.More importantly, the RoBERTa-WWM-BiLSTM-CRF named entity recognition model and RoBERTa-WWM-TextCNN question semantic classification model were constructed to parse the correct semantics of questions from users, which provided algorithm support for the home appliance QA system.The results showed that the method outperformed baselines, and the QA model showed high interpretability and good performance.

Cross-domain recommendation is an effective way to cope with the cold-start problem in recommendation systems. Knowledge of the current, particularly reviews, is taken into account to improve user/item embedding to reduce the neg-ative transfer that occurs during mapping processes across the source and target domains. Traditional approaches, on the other hand, typically apply review information from the source and target domain independently without consideration of user preference divergence. In this paper, we propose a novel Review-based Preference Invariance Transfer Network (RPITN) to minimize negative transfer by combining reviews from two domains. We first build a review preference invari-ance (RPI) embedding procedure to express user/item review correlations between two domains. Then, to improve the gen-eralization ability of user/item embedding and prevent negative transfer across domains, we carefully insert RPI into the embedding learning and mapping process. Extensive exper-iments on real-world datasets demonstrate the superiority of RPITN compared with other recommendation methods.

In this paper, a novel approach based on deep belief networks (DBN) for electrocardiograph (ECG) arrhythmias classification is proposed. The construction process of ECG classification model consists of two steps: features learning for ECG signals and supervised fine-tuning. In order to deeply extract features from continuous ECG signals, two types of restricted Boltzmann machine (RBM) including Gaussian–Bernoulli and Bernoulli–Bernoulli are stacked to form DBN. The parameters of RBM can be learned by two training algorithms such as contrastive divergence and persistent contrastive divergence. A suitable feature representation from the raw ECG data can therefore be extracted in an unsupervised way. In order to enhance the performance of DBN, a fine-tuning process is carried out, which uses backpropagation by adding a softmax regression layer on the top of the resulting hidden representation layer to perform multiclass classification. The method is then validated by experiments on the well-known MIT-BIH arrhythmia database. Considering the real clinical application, the inter-patient heartbeat dataset is divided into two sets and grouped into four classes (N, S, V, F) following the recommendations of AAMI. The experiment results show our approach achieves better performance with less feature learning time than traditional hand-designed methods on the classification of ECG arrhythmias.