Research map / Architectures and efficiency / RNNs and normalization
GroupReduce: Block-Wise Low-Rank Approximation for Neural Language Model Shrinking
Patrick Chen, Si Si, Yang Li, Ciprian Chelba, Cho-Jui Hsieh
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 |
|---|---|---|---|
| Replace standard low-rank approximation with GroupReduce | Neural language model embeddings · OBW | PPL(before retrain): 39.41 -> 32.47 | Improved |
| Replace pruning with GroupReduce | Neural language model embeddings · OBW | PPL(before retrain): 128.31 -> 32.47 | Improved |
| Weight approximation by token frequency (ablation) | Vanilla SVD for PTB-Small embeddings · PTB-Small | PPL: 189.7 -> 179.8 | Improved |
| Use separate low-rank approximations for blocks (ablation) | Weighted low-rank PTB-Small embeddings · PTB-Small | PPL: 179.8 -> 155.3 | Improved |
| Assign more rank to higher-frequency blocks (ablation) | Block low-rank PTB-Small embeddings · PTB-Small | PPL: 155.3 -> 129.2 | Improved |
| Refine block assignments using reconstruction error (ablation) | Dynamic-rank block approximation for PTB-Small · PTB-Small | PPL: 129.2 -> 127.5 | Improved |
| Apply quantization to GroupReduce factors | GroupReduce embeddings · PTB-Small | Embedding Memory compression: 5x -> 40x | Improved |
| Apply quantization to GroupReduce factors | GroupReduce embeddings · OBW | Embedding Memory compression: 6.6x -> 26x | Improved |
| Vary the number of clusters from 5 to 10, 20, or 30 (ablation) | GroupReduce · PTB-Large | PPL(before retrain): 81.79; 80.52; 82.88; 83.1 | Mixed |
| Use GroupReduce rather than deep compositional coding | PTB-Large input embedding · PTB-Large | Embedding Memory compression: 11.8x -> 23.6x | Improved |
| Use GroupReduce rather than deep compositional coding | NMT input embedding · IWSLT 2014 German-to-English | Embedding Memory compression: 16.6x -> 33.3x | Improved |
| Apply low-rank approximation to OBW LSTM cells | OBW-bigLSTM · OBW | PPL | Hurt |
Closest papers by meaning
- Language Models as Zero-shot Lossless Gradient Compressors: Towards General Neural Parameter Prior Models (NeurIPS 2024 · similarity 0.89)
- Efficient Hybrid Language Model Compression through Group-Aware SSM Pruning (NeurIPS 2025 · similarity 0.88)
- Language Modeling Is Compression (ICLR 2024 · similarity 0.88)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.88)
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (NeurIPS 2024 · similarity 0.87)
- Compact Language Models via Pruning and Knowledge Distillation (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.
