Recent years have witnessed an exponential growth of model scale in deep learning-based recommender systems---from Google's 2016 model with 1 billion parameters to the latest Facebook's model with 12 trillion parameters. Significant quality boost has come with each jump of the model capacity, which makes us believe the era of 100 trillion parameters is around the corner. However, the training of such models is challenging even within industrial scale data centers. We resolve this challenge by careful co-design of both optimization algorithm and distributed system architecture. Specifically, to ensure both the training efficiency and the training accuracy, we design a novel hybrid training algorithm, where the embedding layer and the dense neural network are handled by different synchronization mechanisms; then we build a system called Persia (short for parallel recommendation training system with hybrid acceleration) to support this hybrid training algorithm. Both theoretical demonstrations and empirical studies with up to 100 trillion parameters have been conducted to justify the system design and implementation of Persia. We make Persia publicly available (at github.com/PersiaML/Persia) so that anyone can easily train a recommender model at the scale of 100 trillion parameters.

Despite the tremendous success in computer vision, deep convolutional networks suffer from serious computation costs and redundancies. Although previous works address that by enhancing the diversities of filters, they have not considered the complementarity and the completeness of the internal convolutional structure. To respond to this problem, we propose a novel inner-imaging (InI) architecture, which allows relationships between channels to meet the above requirement. Specifically, we organize the channel signal points in groups using convolutional kernels to model both the intragroup and intergroup relationships simultaneously. A convolutional filter is a powerful tool for modeling spatial relations and organizing grouped signals, so the proposed methods map the channel signals onto a pseudoimage, like putting a lens into the internal convolution structure. Consequently, not only is the diversity of channels increased but also the complementarity and completeness can be explicitly enhanced. The proposed architecture is lightweight and easy to be implement. It provides an efficient self-organization strategy for convolutional networks to improve their performance. Extensive experiments are conducted on multiple benchmark datasets, including CIFAR, SVHN, and ImageNet. Experimental results verify the effectiveness of the InI mechanism with the most popular convolutional networks as the backbones.

Physical signs of patients indicate crucial evidence for diagnosing both location and nature of the disease, where there is a sequential relationship between the two tasks. Thus their joint learning can utilize intrinsic association by transferring related knowledge across relevant tasks. Choosing the right time to transfer is a critical problem for joint learning. However, how to dynamically adjust when tasks interact to capture the right time for transferring related knowledge is still an open issue. To this end, we propose a Task-Coupling Elastic Learning (TCEL) framework to model the task relatedness for classifying disease-location and disease-nature based on physical sign images. The main idea is to dynamically transfer relevant knowledge by progressively shifting task-coupling from loose to tight during the multi-stage training. In the early stage of training, we relax the constraints of modeling relations to focus more in learning the generic task-common features. In the later stage, the semantic guidance will be strengthened to learn the task-specific features. Specifically, a dynamic sequential module (DSM) is proposed to explicitly model the sequential relationship and enable multi-stage training. Moreover, to address the side effect of DSM, a new loss regularization is proposed. The extensive experiments on these two clinical datasets show the superiority of the proposed method over the baselines, and demonstrate the effectiveness of the proposed task-coupling elastic mechanism.

Zero-shot learning uses semantic attributes to connect the search space of unseen objects. In recent years, although the deep convolutional network brings powerful visual modeling capabilities to the ZSL task, its visual features have severe pattern inertia and lack of representation of semantic relationships, which leads to severe bias and ambiguity. In response to this, we propose the Graph-based Visual-Semantic Entanglement Network to conduct graph modeling of visual features, which is mapped to semantic attributes by using a knowledge graph, it contains several novel designs: 1. it establishes a multi-path entangled network with the convolutional neural network (CNN) and the graph convolutional network (GCN), which input the visual features from CNN to GCN to model the implicit semantic relations, then GCN feedback the graph modeled information to CNN features; 2. it uses attribute word vectors as the target for the graph semantic modeling of GCN, which forms a self-consistent regression for graph modeling and supervise GCN to learn more personalized attribute relations; 3. it fuses and supplements the hierarchical visual-semantic features refined by graph modeling into visual embedding. Our method outperforms state-of-the-art approaches on multiple representative ZSL datasets: AwA2, CUB, and SUN by promoting the semantic linkage modelling of visual features.

Online healthcare consultation offers people a convenient way to consult doctors. In this paper, we aim at building a generative dialog system for Chinese healthcare consultation. As the original Seq2seq architecture tends to suffer the issue of generating low-quality responses, the multi-source Seq2seq architecture generating more informative responses is much more preferred in this task. The multi-source Seq2seq architecture takes advantage of retrieval techniques to obtain responses from the database, and then takes these responses alongside the user-issued question as input. However, some of the retrieved responses might be not much related to the user-issued question, resulting in the generation of unsatisfying responses that are not correct in diagnosis or instead provide inappropriate advice on prevention or treatment. Therefore, this paper proposes multi-source Seq2seq guided by knowledge (MSSGK) to handle this problem. MSSGK differs from the multi-source Seq2seq architecture in that domain knowledge, including disease labels and topic labels about prevention and treatment, is introduced into the response generation via a multi-task learning framework. To better exploit the domain knowledge, we propose three attention mechanisms to provide more appropriate guidance for response generation. Experimental results on a dataset of real-world healthcare consultation show the effectiveness of the proposed method.

Food recognition plays a much critical role in various health-care applications. However, it poses many challenges to current approaches due to the diverse appearances of food dishes and the non-uniform composition of ingredients for the foods in the same category. Current methods primarily focus on the appearance of foods without considering their semantic information, easily finding the wrong attention areas of food images. Second, these methods lack the dynamic weighting of multiple semantic features in the modeling process. Thus this paper proposes a novel Multi-View Attention Network within the multi-task learning framework that incorporates multiple semantic features into the food recognition task from both ingredient recognition and recipe modeling. It also utilizes the multi-view attention mechanism to automatically adjust the weights of different semantic features and enables different tasks to interact with each other so as to obtain a more comprehensive feature representation. The experiments conducted on both ChineseFoodNet and VIREO Food-172 benchmark databases validate the proposed method with the obvious improvement of the performance and the lower parameter size.