Research map / Architectures and efficiency / RNNs and normalization
Skip RNN: Learning to Skip State Updates in Recurrent Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Learn to skip state updates rather than skip them randomly | LSTM and GRU · MNIST test | Accuracy: 0.893 ± 0.003 -> 0.973 ± 0.002 | Helped |
| Learn to skip state updates rather than skip them randomly | LSTM and GRU · Adding task | Task solved | Helped |
| Add a budget penalty to learned state skipping | Skip LSTM · UCF-101 validation split 1 | Accuracy: 0.671 -> 0.790 | Helped |
| Increase the budget penalty from λ=10⁻⁴ to λ=10⁻² (ablation) | Skip GRU · Charades validation | mAP (%): 8.76 -> 7.95 | Hurt |
| Use learned rather than random state skipping | LSTM · Charades validation | mAP (%): 8.94 -> 8.94 | No clear effect |
| Add learned state skipping with λ=10⁻⁴ | LSTM · IMDB test, length 200 | Accuracy: 0.843 ± 0.003 -> 0.846 ± 0.004 | No clear effect |
| Add learned state skipping with λ=10⁻⁴ | LSTM · IMDB test, length 400 | Accuracy: 0.868 ± 0.004 -> 0.862 ± 0.003 | Hurt |
| Increase the input sequence length and let the model select tokens, rather than crop at length 200 | Skip LSTM · IMDB test | Accuracy: 0.843 ± 0.003 -> 0.862 ± 0.003 | Helped |
| Use a larger budget penalty, λ=10⁻³ rather than λ=10⁻⁴ (ablation) | Skip LSTM · IMDB test, length 200 | Accuracy: 0.837 ± 0.006 -> 0.811 ± 0.007 | Hurt |
| Add learned state skipping without a budget penalty | LSTM · UCF-101 validation split 1 | Accuracy: 0.671 -> 0.749 | Helped |
| Reduce the cost per sample from λ=10⁻⁴ to λ=10⁻⁵ (ablation) | Skip LSTM · UCF-101 validation split 1 | Accuracy: 0.790 -> 0.757 | Mixed |
Closest papers by meaning
- SkipW: Resource Adaptable RNN with Strict Upper Computational Limit (ICLR 2021 · similarity 0.92)
- RNNs Incrementally Evolving on an Equilibrium Manifold: A Panacea for Vanishing and Exploding Gradients? (ICLR 2020 · similarity 0.90)
- Noisin: Unbiased Regularization for Recurrent Neural Networks (ICML 2018 · similarity 0.89)
- Fast-Slow Recurrent Neural Networks (NeurIPS 2017 · similarity 0.89)
- Dilated Recurrent Neural Networks (NeurIPS 2017 · similarity 0.89)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 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.