Researchers have developed a new spatiotemporal Gated Recurrent Unit (ST-GRU) as an alternative to the Spatio-Temporal Long Short-Term Memory (ST-LSTM) unit used in PredRNN-V2. This ST-GRU unit simplifies the recurrent cell by using two GRU-style gates instead of three ST-LSTM gates and merging the temporal memory state into the hidden state, resulting in approximately a 5% reduction in parameters. When integrated into PredRNN-V2 and tested on the Moving MNIST benchmark, the ST-GRU-based model achieved competitive results, including a lower LPIPS score, suggesting its potential as a building block for spatiotemporal predictive learning. AI
IMPACT Introduces a more parameter-efficient recurrent cell for spatiotemporal prediction tasks.
RANK_REASON The cluster contains a research paper detailing a new model architecture. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- gated recurrent unit
- PredRNN-V2
- Recurrent Neural Networks for Predictive Learning using Spatiotemporal LSTMs
- ST-GRU
- ST-GRU PredRNN-V2
- ST-LSTM: A Deep Learning Approach Combined Spatio-Temporal Features for Short-Term Forecast in Rail Transit
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
