Self-Attention ConvLSTM for Spatiotemporal Prediction

Abstract

Spatiotemporal prediction is challenging due to the complex dynamic motion and appearance changes. Existing work concentrates on embedding additional cells into the standard ConvLSTM to memorize spatial appearances during the prediction. These models always rely on the convolution layers to capture the spatial dependence, which are local and inefficient. However, long-range spatial dependencies are significant for spatial applications. To extract spatial features with both global and local dependencies, we introduce the self-attention mechanism into ConvLSTM. Specifically, a novel self-attention memory (SAM) is proposed to memorize features with long-range dependencies in terms of spatial and temporal domains. Based on the self-attention, SAM can produce features by aggregating features across all positions of both the input itself and memory features with pair-wise similarity scores. Moreover, the additional memory is updated by a gating mechanism on aggregated features and an established highway with the memory of the previous time step. Therefore, through SAM, we can extract features with long-range spatiotemporal dependencies. Furthermore, we embed the SAM into a standard ConvLSTM to construct a self-attention ConvLSTM (SA-ConvLSTM) for the spatiotemporal prediction. In experiments, we apply the SA-ConvLSTM to perform frame prediction on the MovingMNIST and KTH datasets and traffic flow prediction on the TexiBJ dataset. Our SA-ConvLSTM achieves state-of-the-art results on both datasets with fewer parameters and higher time efficiency than previous state-of-the-art method.

Polar Vortex Multi-Day Intensity Prediction Relying on New Deep Learning Model: A Combined Convolution Neural Network with Long Short-Term Memory Based on Gaussian Smoothing Method

“…The prediction of polar vortex intensity can be improved by using the method of the benchmark model. The combination of the self-attention mechanism and the ConvLSTM model adopted in spatiotemporal prediction achieves state-of-the-art results [95]. Using the most advanced attention mechanism to predict a certain day or a certain strong/weak event of the polar vortex in a long time series may achieve better results.…”

A Novel Multi-Input Multi-Output Recurrent Neural Network Based on Multimodal Fusion and Spatiotemporal Prediction for 0–4 Hour Precipitation Nowcasting

“…The hidden state of the minor input only depends on its time memory, which is conducive to preserving the spatiotemporal transformation information of the minor input data. In addition, to improve the MFSP-LSTM unit's adaptability to long-distance dependence, we followed Lin's work [27] and added the SAM to the MFSP-LSTM unit. We expanded the global spatiotemporal receptive field of H t and N t by the memory Mt.…”

“…In order to improve the longterm extrapolation ability of the spatiotemporal prediction model, HPRNN [26] proposes a hierarchical prediction strategy which reduces the accumulation of prediction errors over time through a recurrent coarse-to-fine mechanism. Lin et al [27] proposed self-attention memory (SAM) to memorize the features of long-distance dependence in terms of spatial and temporal domains. SAM can be embedded in most spatiotemporal prediction recurrent neural networks.…”

Contextual Sa-Attention Convolutional LSTM for Precipitation Nowcasting: A Spatiotemporal Sequence Forecasting View

“…At the same time, deep learning has also been applied to natural language processing (NLP) [16], [17] and sequence learning [18], [19]. Long short-term memory (LSTM) [20], [21] can effectively learn the inherent relationships of long distance objects. The LSTM encoder-decoder framework [18] effectively solves the sequence-to-sequence learning problems using temporally concatenated LSTM.…”