Scholarlink

<p>The core challenge in spati-temporal prediction learning lies in effectively and collaboratively capturing global spatial dependencies and long-term temporal dynamics. Existing methods often face a trade-off: convolutional-based recurrent networks (e.g., ConvLSTM), while proficient in sequential modeling, are limited by local receptive fields, hindering their ability to capture long-range spatial correlations. Conversely, pure Transformer architectures, despite their strong global modeling capability, may weaken the inherent advantage of recurrent networks in handling fine-grained temporal continuity. To address this dilemma, we propose a novel recurrent prediction architecture, GAttenRNN. Our main contribution is the design of an innovative recurrent unit, AttenLSTM, which, embeds a gated Transformer module as the core feature extractor within the recurrent network framework, replacing traditional convolutional operations. This design aims to deeply fuse the powerful global context modeling ability of the Transformer with the sequential processing capability of the recurrent neural network. Furthermore, our architecture incorporates a dual memory stream mechanism to enhance dynamic information interaction across different layers. Experimental results on three benchmark datasets—moving MNIST, TaxiBJ, and KTH—demonstrate that GAttenRNN, as a concise and highly efficient framework, achieves highly competitive prediction accuracy, fully validating the effectiveness of our proposed fusion architecture. The code and models are publicly available at <a href="https://github.com/wanwan-del/GAttenRNN">https://github.com/wanwan-del/GAttenRNN</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GAttenRNN: a recurrent neural network for spati-temporal prediction learning based on gated transformer

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Gaihui Guo,
  • Jiaxin Wan,
  • Weichuan Zhang

摘要

The core challenge in spati-temporal prediction learning lies in effectively and collaboratively capturing global spatial dependencies and long-term temporal dynamics. Existing methods often face a trade-off: convolutional-based recurrent networks (e.g., ConvLSTM), while proficient in sequential modeling, are limited by local receptive fields, hindering their ability to capture long-range spatial correlations. Conversely, pure Transformer architectures, despite their strong global modeling capability, may weaken the inherent advantage of recurrent networks in handling fine-grained temporal continuity. To address this dilemma, we propose a novel recurrent prediction architecture, GAttenRNN. Our main contribution is the design of an innovative recurrent unit, AttenLSTM, which, embeds a gated Transformer module as the core feature extractor within the recurrent network framework, replacing traditional convolutional operations. This design aims to deeply fuse the powerful global context modeling ability of the Transformer with the sequential processing capability of the recurrent neural network. Furthermore, our architecture incorporates a dual memory stream mechanism to enhance dynamic information interaction across different layers. Experimental results on three benchmark datasets—moving MNIST, TaxiBJ, and KTH—demonstrate that GAttenRNN, as a concise and highly efficient framework, achieves highly competitive prediction accuracy, fully validating the effectiveness of our proposed fusion architecture. The code and models are publicly available at https://github.com/wanwan-del/GAttenRNN.