Research map / Architectures and efficiency / RNNs and normalization
PredRNN: A Recurrent Neural Network for Spatiotemporal Predictive Learning
Yunbo Wang, Haixu Wu, Jianjin Zhang, Zhifeng Gao, Jianmin Wang, Philip S. Yu, Mingsheng Long
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add spatiotemporal memory flow to ConvLSTM (ablation) | ConvLSTM · Moving MNIST (fixed 10k training set) | MSE: 103.3 -> 74.0 | Improved |
| replace ConvLSTM unit with ST-LSTM (ablation) | PredRNN-M-Only · Moving MNIST (fixed 10k training set) | MSE: 74.0 -> 56.8 | Improved |
| add memory decoupling and reverse scheduled sampling (ablation) | PredRNN · Moving MNIST (fixed 10k training set) | MSE: 56.8 -> 48.4 | Improved |
| add memory decoupling alone (ablation) | PredRNN · Moving MNIST (fixed 10k training set) | MSE: 56.8 -> 51.1 | Improved |
| keep full bottom-up memory flow instead of truncating (ablation) | PredRNN · Moving MNIST (fixed 10k training set) | MSE: 57.6 -> 56.8 | Improved |
| keep full top-down memory flow instead of truncating (ablation) | PredRNN · Moving MNIST (fixed 10k training set) | MSE: 59.7 -> 56.8 | Improved |
| add reverse scheduled sampling to ConvLSTM (ablation) | ConvLSTM · Moving MNIST (fixed 10k training set) | MSE: 103.3 -> 64.2 | Improved |
| use reverse scheduled sampling (2nd strategy, exponential) instead of scheduled sampling (ablation) | PredRNN · Moving MNIST (fixed 10k training set) | MSE: 56.8 -> 50.6 | Improved |
| use sigmoid instead of linear/exponential schedule in 1st RSS strategy (ablation) | PredRNN · Moving MNIST (fixed 10k training set) | MSE: 51.8 -> 53.9 | Hurt |
| add decoupling and RSS to PredRNN++ (ablation) | PredRNN++ · Moving MNIST (fixed 10k training set) | MSE: 46.5 -> 44.8 | Improved |
| add decoupling and RSS to E3D-LSTM (ablation) | E3D-LSTM · Moving MNIST (fixed 10k training set) | MSE: 44.2 -> 42.1 | Improved |
| add decoupling and RSS to MIM (ablation) | MIM · Moving MNIST (fixed 10k training set) | MSE: 52.0 -> 45.8 | Improved |
Closest papers by meaning
- PredRNN: Recurrent Neural Networks for Predictive Learning using Spatiotemporal LSTMs (NeurIPS 2017 · similarity 0.95)
- PredRNN++: Towards A Resolution of the Deep-in-Time Dilemma in Spatiotemporal Predictive Learning (ICML 2018 · similarity 0.92)
- ModeRNN: Harnessing Spatiotemporal Mode Collapse in Unsupervised Predictive Learning (TPAMI 2023 · similarity 0.89)
- Making Convolutional Networks Recurrent for Visual Sequence Learning (CVPR 2018 · similarity 0.89)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · similarity 0.88)
- Wavelet-Driven Spatiotemporal Predictive Learning: Bridging Frequency and Time Variations (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
