Research map / Architectures and efficiency / RNNs and normalization
LightRNN: Memory and Computation-Efficient 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 |
|---|---|---|---|
| Use 2-Component shared embeddings instead of conventional independent word embeddings | LightRNN · ACLW datasets | PPL: 157 vs 169 (Spanish); 176 vs 190 (French); 191 vs 216 (English); 558 vs 578 (Czech); 281 vs 305 (German); 288 vs 313 (Ru | Helped |
| Use 2-Component shared embeddings | LightRNN · BillionW | PPL: 66 vs 68 | Helped |
| Increase the shared embedding size from 200 to 600 (ablation) | LightRNN · ACLW-French test | PPL: 340 -> 208 | Helped |
| Increase the shared embedding size from 600 to 1000 (ablation) | LightRNN · ACLW-French test | PPL: 208 -> 176 | Helped |
| Ensemble LightRNN with the KN 5-gram model | LightRNN · BillionW | PPL: 66 -> 43 | Helped |
| Use LightRNN rather than C-HSM | ACLW language modeling · ACLW | runtime (hours) to reach baseline PPL: 168 -> 82 | Helped |
| Use LightRNN rather than HSM | BillionW language modeling · BillionW | runtime (hours) to reach baseline PPL: 168 -> 70 | Helped |
| Refine the word-allocation table over successive bootstrap rounds (ablation) | LightRNN · ACLW language modeling | PPL | Helped |
Closest papers by meaning
- Fast-Slow Recurrent Neural Networks (NeurIPS 2017 · similarity 0.90)
- Tying Word Vectors and Word Classifiers: A Loss Framework for Language Modeling (ICLR 2017 · similarity 0.90)
- FlashRNN: Unlocking Parallel Training of Nonlinear RNNs for Large Language Models (ICLR 2026 · similarity 0.90)
- Quasi-Recurrent Neural Networks (ICLR 2017 · similarity 0.89)
- A Tensor Decomposition Perspective on Second-order RNNs (ICML 2024 · similarity 0.89)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.88)
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.