Encoder-decoder models have been widely used in image captioning, and most of them are designed via single Long Short Term Memory (LSTM). The capacity of single-layer network, whose encoder and decoder are integrated together, is limited for such a complex task of image captioning. Moreover, how to effectively increase the “vertical depth” of encoder-decoder remains to be solved. To deal with these problems, a novel Deep Hierarchical Encoder-Decoder Network (DHEDN) is proposed for image captioning, where a deep hierarchical structure is explored to separate the functions of encoder and decoder. This model is capable of efficiently exerting the representation capacity of deep networks to fuse high level semantics of vision and language in generating captions. Specifically, visual representations in top levels of abstraction are simultaneously considered, and each of these levels is associated to one LSTM. The bottom-most LSTM is applied as the encoder of textual inputs. The application of the middle layer in encoder-decoder is to enhance the decoding ability of top-most LSTM. Furthermore, depending on the introduction of semantic enhancement module of image feature and distribution combine module of text feature, variants of architectures of our model are constructed to explore the impacts and mutual interactions among the visual representation, textual representations and the output of the middle LSTM layer. Particularly, the framework is training under a reinforcement learning method to address the exposure bias problem between the training and the testing by the policy gradient optimization. Qualitative analyses indicate the process that our model “translates” image to sentence and further visualization presents the evolution of the hidden states from different hierarchical LSTMs over time. Extensive experiments demonstrate that our model outperforms current state-of-the-art models on three benchmark datasets: Flickr8K, Flickr30K and MSCOCO. On both image captioning and retrieval tasks, our method achieves the best results. On MSCOCO captioning Leaderboard, our method also achieves superior performance.

Precipitation plays a significant role in global water and energy cycles, largely affecting many aspects of human life, such as transportation and agriculture. Recently, meteorologists have tried to predict precipitation with deep learning methods by learning from much historical meteorological data. Under this paradigm, the task of precipitation nowcasting is formulated as a spatiotemporal sequence forecasting problem. However, current studies suffer from two inherent drawbacks of the definition of the problem. First, considering that the weather patterns vary in spatial and temporal dimensions, a spatiotemporally shared kernel is not optimal for capturing features across different regions and seasons. Second, these methods isolate the precipitation from other meteorological elements, such as temperature, humidity, and wind. The disability of cross-model learning prevents the possibility of the promotion of precipitation prediction. Therefore, this article proposes a spatiotemporal inference network (STIN) to produce precipitation prediction from multimodal meteorological data with spatiotemporal specific filters. Specifically, we first design a spatiotemporal-aware convolutional layer (STAConv), in which kernels are generated conditioned on the incoming spatiotemporally features vector. Replacing normal convolution with STAConv enables the extraction of spatiotemporal specific information from the meteorological data. Based on the STAConv, the spatiotemporal-aware convolutional neural network (STACNN) is further proposed, fusing the multimodal information, including temperature, humidity, and wind. Then, an encoder–decoder framework composed of RNN layers is built to extract representative temporal dynamics from multimodal information. To investigate the practicality of the proposed method, we employ STIN to predict the following precipitation intensity. Extensive experiments on three meteorological datasets demonstrate the effectiveness of our model on precipitation nowcasting.

The primary objective of precipitation nowcasting is to predict precipitation patterns several hours in advance. Recent studies have emphasized the potential of deep learning methods for this task. To harness the correlations among various meteorological elements, existing frameworks project multiple meteorological elements into a latent space and then utilize convolutional-recurrent networks for future precipitation prediction. Although effective, the escalating model complexity may impede practical applications. This paper develops the Preformer, a streamlined Transformer framework for precipitation nowcasting that efficiently captures global spatiotemporal dependencies among multiple meteorological elements. The Preformer implements an encoder-translator-decoder architecture, where the encoder integrates spatial features of multiple elements, the translator models spatiotemporal dynamics, and the decoder combines spatiotemporal information to forecast future precipitation. Without introducing complex structures or strategies, the Preformer achieves state-of-the-art performance even with the least parameters.

Xinyu Xiao, Lingfeng Wang, Bin Fan, Shinming Xiang, Chunhong Pan. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Recently, image attributes containing high-level semantic information have been widely used in computer vision tasks, including visual recognition and image captioning. Existing attribute extraction methods map visual concepts to the probabilities of frequently-used words by directly using Convolutional Neural Networks (CNNs). Typically, two main problems exist in those methods. First, words of different parts of speech (POSs) are handled in the same way, but non-nominal words can hardly be mapped to visual regions through CNNs only. Second, synonymous nominal words are treated as independent and different words, in which similarities are ignored. In this paper, a novel Refined Universal Detection (RUDet) method is proposed to solve these two problems. Specifically, a Refinement (RF) module is designed to extract refined attributes of non-nominal words based on the attributes of nominal words and visual features. In addition, a Word Tree (WT) module is constructed to integrate synonymous nouns, which ensures that similar words hold similar and more accurate probabilities. Moreover, a Feature Enhancement (FE) module is adopted to enhance the ability to mine different visual concepts in different scales. Experiments conducted on the large-scale Microsoft (MS) COCO dataset illustrate the effectiveness of our proposed method.