Research map / Architectures and efficiency / RNNs and normalization
Reversible Recurrent Neural Networks
Matthew MacKay, Paul Vicol, Jimmy Ba, Roger B Grosse
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 |
|---|---|---|---|
| Use reversible recurrent transitions with information buffering | GRU language models on Penn TreeBank · Penn TreeBank validation | perplexity: 82.2 -> 81.5 | No clear effect |
| Use reversible recurrent transitions with information buffering | LSTM language models on WikiText-2 · WikiText-2 validation | perplexity: 89.3 -> 94.1 | Hurt |
| Use Emb+20H attention and reversible recurrence | GRU translation model · Multi30K test | BLEU: 32.60 -> 34.41 | Improved |
| Use Emb+20H attention and reversible recurrence | LSTM translation model · Multi30K test | BLEU: 37.06 -> 37.23 | Improved |
| Use reversible recurrence | GRU translation model · IWSLT 2016 test | BLEU: 16.07 -> 20.70 | Improved |
| Use reversible recurrence | LSTM translation model · IWSLT 2016 test | BLEU: 22.35 -> 22.34 | No clear effect |
| Restrict forgetting to 2–3 bits per hidden unit per timestep (ablation) | RevGRU and RevLSTM language models · Penn TreeBank validation | perplexity | No clear effect |
| Restrict forgetting to at most 1 bit per hidden unit per timestep (ablation) | RevGRU language model · Penn TreeBank validation | perplexity: 4–5 increase in perplexity | Hurt |
| Use discrete instead of fractional forgetting (ablation) | Reversible language models · Penn TreeBank validation | perplexity: 4-10 perplexity points | Hurt |
| Use straight-through Gumbel-Softmax instead of straight-through categorical gradient estimation (ablation) | Discrete-forgetting models · Penn TreeBank validation | perplexity: 2--3 perplexity | Improved |
| Use attention over embeddings and a 20-dimensional hidden-state slice instead of other attention variants (ablation) | Reversible NMT models · Multi30K test | BLEU | Mixed |
| Remove the forgetting step (ablation) | Reversible recurrent models · Penn TreeBank language modeling | perplexity | Hurt |
Closest papers by meaning
- Capacity and Trainability in Recurrent Neural Networks (ICLR 2017 · similarity 0.88)
- Time Adaptive Recurrent Neural Network (CVPR 2021 · similarity 0.87)
- Improved memory in recurrent neural networks with sequential non-normal dynamics (ICLR 2020 · similarity 0.87)
- Skip RNN: Learning to Skip State Updates in Recurrent Neural Networks (ICLR 2018 · similarity 0.87)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.86)
- Memory-Efficient Reversible Spiking Neural Networks (AAAI 2024 · 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.
