Research map / Architectures and efficiency / RNNs and normalization
Learning Sequence Representations by Non-local Recurrent Neural Memory
Wenjie Pei, Xin Feng, Canmiao Fu, Qiong Cao, Guangming Lu, Yu‐Wing Tai
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 NRNM memory module to LSTM backbone | LSTM · NTU RGB+D Cross-View | accuracy: 84.0 -> 89.9 | Improved |
| add NRNM memory module to LSTM backbone | LSTM · NTU RGB+D Cross-Subject | accuracy: 70.3 -> 80.8 | Improved |
| add NRNM to LSTM on I3D features | LSTM · Charades (R101-I3D features) | mAP: 35.7 -> 40.1 | Improved |
| add NRNM to LSTM on I3D-NL features | LSTM · Charades (R101-I3D-NL features) | mAP: 38.9 -> 41.7 | Improved |
| use self-attention instead of concatenation plus linear transform for non-local operations in NRNM (ablation) | NRNM · NTU RGB+D | accuracy | Improved |
| replace NRNM by a tuned Transformer (ablation) | NRNM vs Transformer · NTU RGB+D Cross-View | accuracy: 89.9 -> 81.3 | Hurt |
| increase LSTM depth and width to match NRNM capacity (ablation) | LSTM · NTU RGB+D Cross-View | accuracy: 83.9 -> 84.0 | No clear effect |
| deeper EleAtt-GRU (5 layers vs 3 layers, 256 units) (ablation) | EleAtt-GRU · NTU RGB+D Cross-View | accuracy: 85.4 -> 85.0 | Hurt |
| integrate NRNM at one layer versus several layers of LSTM backbone (ablation) | NRNM · NTU RGB+D Cross-View | accuracy | Improved |
| increase NRNM block size beyond 8 (ablation) | NRNM · NTU RGB+D Cross-View | accuracy | Hurt |
| set NRNM hidden units to 512 versus fewer or more (ablation) | NRNM · NTU RGB+D Cross-View | accuracy | Improved |
| use RGB-based LFB instead of skeleton-based NRNM | NRNM vs LFB · NTU RGB+D Cross-View | accuracy: 89.9 -> 93.4 | Mixed |
Closest papers by meaning
- Non-Local Recurrent Neural Memory for Supervised Sequence Modeling (ICCV 2019 · similarity 0.95)
- Kernel-Based Approaches for Sequence Modeling: Connections to Neural Methods (NeurIPS 2019 · similarity 0.88)
- State-Frequency Memory Recurrent Neural Networks (ICML 2017 · similarity 0.88)
- Long Expressive Memory for Sequence Modeling (ICLR 2022 · similarity 0.87)
- Towards Non-Saturating Recurrent Units for Modelling Long-Term Dependencies (AAAI 2019 · similarity 0.87)
- RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling (NeurIPS 2025 · similarity 0.86)
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.
