Research map / Architectures and efficiency / RNNs and normalization
Low-Rank Constraints for Fast Inference in Structured Models
Justin T Chiu, Yuntian Deng, Alexander M Rush
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 |
|---|---|---|---|
| impose low-rank constraint (8:1 state-to-rank) on HMM transitions | HMM, 2^14 states · PTB val | perplexity: 144.3 -> 141.4 | Improved |
| impose low-rank constraint (8:1) on HMM transitions | HMM, 2^14 states · PTB test | perplexity: 136.8 -> 131.8 | Improved |
| raise state-to-rank ratio from 8 to 32 (ablation) | LHMM · PTB val | perplexity: 141.4 -> 153.7 | Hurt |
| add banded high-rank transition component at ratio 16 (ablation) | LHMM · PTB val | perplexity: 146.3 -> 138.8 | Improved |
| add banded component at ratio 32 (ablation) | LHMM · PTB val | perplexity: 153.7 -> 145.0 | Improved |
| add banded component at ratio 8 (ablation) | LHMM · PTB val | perplexity: 141.4 -> 143.8 | Hurt |
| use block-sparse emission constraint (VL-HMM) instead of plain HMM | HMM · PTB test | perplexity: 136.8 -> 119.5 | Improved |
| state-to-rank ratio 8:1 at 2^14 states for speed | LHMM · PTB | seconds per batch: more than 3x speedup | Improved |
| low-rank constraint at small HMM sizes (below 2^12) | LHMM · PTB val | perplexity | Hurt |
| low-rank constraint on HMM for polyphonic music | LHMM · Polyphonic music (Nott, Piano, Muse, JSB) | NLL | Mixed |
| low-rank PCFG, 30 nonterminals, N=8 | PCFG · PTB test | perplexity: 252.60 -> 247.02 | Improved |
| low-rank PCFG, 60 nonterminals, N=32 | PCFG · PTB test | perplexity: 234.01 -> 213.81 | Improved |
Closest papers by meaning
- Sequences of Logits Reveal the Low Rank Structure of Language Models (ICLR 2026 · similarity 0.89)
- Deep networks learn to parse uniform-depth context-free languages from local statistics (ICML 2026 · similarity 0.87)
- Pivoting Factorization: A Compact Meta Low-Rank Representation of Sparsity for Efficient Inference in Large Language Models (ICML 2025 · similarity 0.86)
- Pre-trained Large Language Models Learn to Predict Hidden Markov Models In-context (NeurIPS 2025 · similarity 0.85)
- Towards a theory of how the structure of language is acquired by deep neural networks (NeurIPS 2024 · similarity 0.85)
- MaskLLM: Learnable Semi-Structured Sparsity for Large Language Models (NeurIPS 2024 · similarity 0.85)
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.
