Recent advances in cross-lingual text-to-speech (TTS) made it possible to synthesize speech in a language foreign to a monolingual speaker. However, there is still a large gap between the pronunciation of generated cross-lingual speech and that of native speakers in terms of naturalness and intelligibility. In this paper, a triplet training scheme is proposed to enhance the cross-lingual pronunciation by allowing previously unseen content and speaker combinations to be seen during training. Proposed method introduces an extra fine-tune stage with triplet loss during training, which efficiently draws the pronunciation of the synthesized foreign speech closer to those from the native anchor speaker, while preserving the non-native speaker's timbre. Experiments are conducted based on a state-of-the-art baseline cross-lingual TTS system and its enhanced variants. All the objective and subjective evaluations show the proposed method brings significant improvement in both intelligibility and naturalness of the synthesized cross-lingual speech.
Hard interaction learning between source sequences and their next targets is challenging, which exists in a myriad of sequential prediction tasks. During the training process, most existing methods focus on explicitly hard interactions caused by wrong responses. However, a model might conduct correct responses by capturing a subset of learnable patterns, which results in implicitly hard interactions with some unlearned patterns. As such, its generalization performance is weakened. The problem gets more serious in sequential prediction due to the interference of substantial similar candidate targets. To this end, we propose a Hardness Aware Interaction Learning framework (HAIL) that mainly consists of two base sequential learning networks and mutual exclusivity distillation (MED). The base networks are initialized differently to learn distinctive view patterns, thus gaining different training experiences. The experiences in the form of the unlikelihood of correct responses are drawn from each other by MED, which provides mutual exclusivity knowledge to figure out implicitly hard interactions. Moreover, we deduce that the unlikelihood essentially introduces additional gradients to push the pattern learning of correct responses. Our framework can be easily extended to more peer base networks. Evaluation is conducted on four datasets covering cyber and physical spaces. The experimental results demonstrate that our framework outperforms several state-of-the-art methods in terms of top-k based metrics.
Unsupervised graph representation learning has emerged as a powerful tool to address real-world problems and achieves huge success in the graph learning domain. Graph contrastive learning is one of the unsupervised graph representation learning methods, which recently attracts attention from researchers and has achieved state-of-the-art performances on various tasks. The key to the success of graph contrastive learning is to construct proper contrasting pairs to acquire the underlying structural semantics of the graph. However, this key part is not fully explored currently, most of the ways generating contrasting pairs focus on augmenting or perturbating graph structures to obtain different views of the input graph. But such strategies could degrade the performances via adding noise into the graph, which may narrow down the field of the applications of graph contrastive learning. In this paper, we propose a novel graph contrastive learning method, namely \textbf{D}ual \textbf{S}pace \textbf{G}raph \textbf{C}ontrastive (DSGC) Learning, to conduct graph contrastive learning among views generated in different spaces including the hyperbolic space and the Euclidean space. Since both spaces have their own advantages to represent graph data in the embedding spaces, we hope to utilize graph contrastive learning to bridge the spaces and leverage advantages from both sides. The comparison experiment results show that DSGC achieves competitive or better performances among all the datasets. In addition, we conduct extensive experiments to analyze the impact of different graph encoders on DSGC, giving insights about how to better leverage the advantages of contrastive learning between different spaces.
Deep learning in molecular and materials sciences is limited by the lack of integration between applied science, artificial intelligence, and high-performance computing. Bottlenecks with respect to the amount of training data, the size and complexity of model architectures, and the scale of the compute infrastructure are all key factors limiting the scaling of deep learning for molecules and materials. Here, we present $\textit{LitMatter}$, a lightweight framework for scaling molecular deep learning methods. We train four graph neural network architectures on over 400 GPUs and investigate the scaling behavior of these methods. Depending on the model architecture, training time speedups up to $60\times$ are seen. Empirical neural scaling relations quantify the model-dependent scaling and enable optimal compute resource allocation and the identification of scalable molecular geometric deep learning model implementations.
Recent advances in path-based explainable recommendation systems have attracted increasing attention thanks to the rich information provided by knowledge graphs. Most existing explainable recommendations only utilize static knowledge graphs and ignore the dynamic user-item evolutions, leading to less convincing and inaccurate explanations. Although there are some works that realize that modelling user's temporal sequential behaviour could boost the performance and explainability of the recommender systems, most of them either only focus on modelling user's sequential interactions within a path or independently and separately of the recommendation mechanism. In this paper, we propose a novel Temporal Meta-path Guided Explainable Recommendation leveraging Reinforcement Learning (TMER-RL), which utilizes reinforcement item-item path modelling between consecutive items with attention mechanisms to sequentially model dynamic user-item evolutions on dynamic knowledge graph for explainable recommendation. Compared with existing works that use heavy recurrent neural networks to model temporal information, we propose simple but effective neural networks to capture users' historical item features and path-based context to characterize the next purchased item. Extensive evaluations of TMER on two real-world datasets show state-of-the-art performance compared against recent strong baselines.
This paper introduces a two-phase deep feature engineering framework for efficient learning of semantics enhanced joint embedding, which clearly separates the deep feature engineering in data preprocessing from training the text-image joint embedding model. We use the Recipe1M dataset for the technical description and empirical validation. In preprocessing, we perform deep feature engineering by combining deep feature engineering with semantic context features derived from raw text-image input data. We leverage LSTM to identify key terms, deep NLP models from the BERT family, TextRank, or TF-IDF to produce ranking scores for key terms before generating the vector representation for each key term by using word2vec. We leverage wideResNet50 and word2vec to extract and encode the image category semantics of food images to help semantic alignment of the learned recipe and image embeddings in the joint latent space. In joint embedding learning, we perform deep feature engineering by optimizing the batch-hard triplet loss function with soft-margin and double negative sampling, taking into account also the category-based alignment loss and discriminator-based alignment loss. Extensive experiments demonstrate that our SEJE approach with deep feature engineering significantly outperforms the state-of-the-art approaches.
Video captioning is a challenging task that captures different visual parts and describes them in sentences, for it requires visual and linguistic coherence. The attention mechanism in the current video captioning method learns to assign weight to each frame, promoting the decoder dynamically. This may not explicitly model the correlation and the temporal coherence of the visual features extracted in the sequence frames.To generate semantically coherent sentences, we propose a new Visual-aware Attention (VA) model, which concatenates dynamic changes of temporal sequence frames with the words at the previous moment, as the input of attention mechanism to extract sequence features.In addition, the prevalent approaches widely use the teacher-forcing (TF) learning during training, where the next token is generated conditioned on the previous ground-truth tokens. The semantic information in the previously generated tokens is lost. Therefore, we design a self-forcing (SF) stream that takes the semantic information in the probability distribution of the previous token as input to enhance the current token.The Dual-stream Decoder (DD) architecture unifies the TF and SF streams, generating sentences to promote the annotated captioning for both streams.Meanwhile, with the Dual-stream Decoder utilized, the exposure bias problem is alleviated, caused by the discrepancy between the training and testing in the TF learning.The effectiveness of the proposed Visual-aware Attention Dual-stream Decoder (VADD) is demonstrated through the result of experimental studies on Microsoft video description (MSVD) corpus and MSR-Video to text (MSR-VTT) datasets.
Multi-object tracking and segmentation (MOTS) is a critical task for autonomous driving applications. The existing MOTS studies face two critical challenges: 1) the published datasets inadequately capture the real-world complexity for network training to address various driving settings; 2) the working pipeline annotation tool is under-studied in the literature to improve the quality of MOTS learning examples. In this work, we introduce the DG-Labeler and DGL-MOTS dataset to facilitate the training data annotation for the MOTS task and accordingly improve network training accuracy and efficiency. DG-Labeler uses the novel Depth-Granularity Module to depict the instance spatial relations and produce fine-grained instance masks. Annotated by DG-Labeler, our DGL-MOTS dataset exceeds the prior effort (i.e., KITTI MOTS and BDD100K) in data diversity, annotation quality, and temporal representations. Results on extensive cross-dataset evaluations indicate significant performance improvements for several state-of-the-art methods trained on our DGL-MOTS dataset. We believe our DGL-MOTS Dataset and DG-Labeler hold the valuable potential to boost the visual perception of future transportation.
Math word problem (MWP) is a challenging and critical task in natural language processing. Many recent studies formalize MWP as a generation task and have adopted sequence-to-sequence models to transform problem descriptions to mathematical expressions. However, mathematical expressions are prone to minor mistakes while the generation objective does not explicitly handle such mistakes. To address this limitation, we devise a new ranking task for MWP and propose Generate & Rank, a multi-task framework based on a generative pre-trained language model. By joint training with generation and ranking, the model learns from its own mistakes and is able to distinguish between correct and incorrect expressions. Meanwhile, we perform tree-based disturbance specially designed for MWP and an online update to boost the ranker. We demonstrate the effectiveness of our proposed method on the benchmark and the results show that our method consistently outperforms baselines in all datasets. Particularly, in the classical Math23k, our method is 7% (78.4% $\rightarrow$ 85.4%) higher than the state-of-the-art.
User purchasing prediction with multi-behavior information remains a challenging problem for current recommendation systems. Various methods have been proposed to address it via leveraging the advantages of graph neural networks (GNNs) or multi-task learning. However, most existing works do not take the complex dependencies among different behaviors of users into consideration. They utilize simple and fixed schemes, like neighborhood information aggregation or mathematical calculation of vectors, to fuse the embeddings of different user behaviors to obtain a unified embedding to represent a user's behavioral patterns which will be used in downstream recommendation tasks. To tackle the challenge, in this paper, we first propose the concept of hyper meta-path to construct hyper meta-paths or hyper meta-graphs to explicitly illustrate the dependencies among different behaviors of a user. How to obtain a unified embedding for a user from hyper meta-paths and avoid the previously mentioned limitations simultaneously is critical. Thanks to the recent success of graph contrastive learning, we leverage it to learn embeddings of user behavior patterns adaptively instead of assigning a fixed scheme to understand the dependencies among different behaviors. A new graph contrastive learning based framework is proposed by coupling with hyper meta-paths, namely HMG-CR, which consistently and significantly outperforms all baselines in extensive comparison experiments.