Research map / Time series / Causal and generative time series
Preventing Gradient Explosions in Gated Recurrent Units
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 |
|---|---|---|---|
| replace gradient clipping (threshold 10) with spectral norm constraint (delta 0.2) | GRU · Penn Treebank language modeling | test perplexity: 100.4 -> 97.6 | Helped |
| replace gradient clipping (threshold 10) with spectral norm constraint | GRU · Penn Treebank language modeling | training success rate (%): 40 -> 100 | Helped |
| replace gradient clipping (threshold 15) with spectral norm constraint | GRU · Nottingham polyphonic music | test negative log-likelihood: 3.64 -> 3.53 | Helped |
| increase clipping threshold from 5 to 10 (ablation) | GRU with gradient clipping · Penn Treebank language modeling | test perplexity: 106.9 -> 100.4 | Helped |
| increase delta from 0.2 to 1.4 (ablation) | Spectral-norm GRU · Penn Treebank language modeling | test perplexity: 97.6 -> 102.1 | Hurt |
| increase delta from 0.5 to 1.1 or 1.4 (ablation) | Spectral-norm GRU · Nottingham polyphonic music | test negative log-likelihood: 3.53 -> 4.56 | Hurt |
| use truncated SVD instead of naive SVD for projection | Spectral-norm GRU · Penn Treebank language modeling | computation time (s): 5.02 × 104 -> 4.55 × 104 | Helped |
| replace gradient clipping with spectral norm constraint, truncated SVD | GRU · Penn Treebank language modeling | computation time (s): 4.96 × 104 -> 4.55 × 104 | No clear effect |
| increase gradient clipping threshold from 15 to 60 (ablation) | GRU with gradient clipping · Nottingham polyphonic music | test negative log-likelihood: 3.64 -> 5.36 | Hurt |
Closest papers by meaning
- The Statistical Recurrent Unit (ICML 2017 · similarity 0.88)
- A recurrent neural network without chaos (ICLR 2017 · similarity 0.87)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.87)
- Can recurrent neural networks warp time? (ICLR 2018 · similarity 0.86)
- Parallel Training of GRU Networks with a Multi-Grid Solver for Long Sequences (ICLR 2022 · similarity 0.86)
- Parallelizing non-linear sequential models over the sequence length (ICLR 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.