Research map / Architectures and efficiency / RNNs and normalization
Learning Long-Term Dependencies for Action Recognition With a Biologically-Inspired Deep Network
Yemin Shi, Yonghong Tian, Yaowei Wang, Wei Zeng, Tiejun Huang
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 |
|---|---|---|---|
| replace 2-layer GRU with shuttleNet (2 processors, 2 steps) | CNN-RNN two-stream on GoogLeNet · UCF101 split 1 (GoogLeNet base) | two-stream accuracy: 90.27% -> 91.94% | Improved |
| replace 2-layer LSTM with shuttleNet | CNN-RNN two-stream on GoogLeNet · UCF101 split 1 (GoogLeNet base) | two-stream accuracy: 88.58% -> 91.94% | Improved |
| add RNN (GRU, 2 layers) to GoogLeNet | GoogLeNet · UCF101 split 1 (GoogLeNet base) | two-stream accuracy: 85.91% -> 90.27% | Improved |
| add output selector (attention over 2 processors) in 1 step | shuttleNet 2 processors 1 step · UCF101 split 1 two streams | two-stream accuracy: 91.7% vs 90.27% (2-layer GRU) | Improved |
| share processor across 2 steps (loop connection) with 1 processor (ablation) | shuttleNet · UCF101 split 1 two streams | two-stream accuracy: 90.27% -> 91.65% | Improved |
| use 2 steps instead of 1 step (ablation) | shuttleNet 2 processors · UCF101 split 1 spatial stream, shuttleNet setting study | spatial accuracy: 82.37% -> 83.03% | Improved |
| use 3 steps instead of 2 steps (ablation) | shuttleNet 2 processors · UCF101 split 1 spatial stream, shuttleNet setting study | spatial accuracy: 83.03% -> 80.89% | Hurt |
| increase processors from 1 to 3 at 2 steps (ablation) | shuttleNet · UCF101 split 1 spatial stream, shuttleNet setting study | spatial accuracy: 82.9% -> 83.19% | Improved |
| increase processors and steps to 3 and 3 (ablation) | shuttleNet · UCF101 split 1 spatial stream, shuttleNet setting study | spatial accuracy: 83.03% -> 82.13% | Hurt |
| switch base CNN from GoogLeNet to Inception-ResNet-v2 | shuttleNet · UCF101 (mean of 3 splits) | two-stream accuracy: 92.3% -> 94.4% | Improved |
| late-fuse hand-crafted MIFS features | shuttleNet two streams · HMDB51 (mean of 3 splits) | accuracy: 66.6% -> 71.7% | Improved |
| late-fuse MIFS features | shuttleNet two streams · UCF101 (mean of 3 splits) | accuracy: 94.4% -> 95.4% | Improved |
Closest papers by meaning
- Training biologically plausible recurrent neural networks on cognitive tasks with long-term dependencies (NeurIPS 2023 · similarity 0.89)
- Resurrecting Recurrent Neural Networks for Long Sequences (ICML 2023 · similarity 0.86)
- Temporal Pyramid Recurrent Neural Network (AAAI 2020 · similarity 0.86)
- Hierarchically Gated Recurrent Neural Network for Sequence Modeling (NeurIPS 2023 · similarity 0.86)
- Recurrent neural network dynamical systems for biological vision (NeurIPS 2024 · similarity 0.85)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.84)
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.
