Research map / Architectures and efficiency / Pruning and efficient tuning
ROSITA: Refined BERT cOmpreSsion with InTegrAted techniques
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 a deeper, narrower architecture (Architecture C) rather than a shallower, wider architecture (Architecture I) (ablation) | BERT compression with one-step pruning and KD · GLUE dev subset | Avg: 75.4 -> 78.6 | Helped |
| Allocate more compression to intermediate outputs and embedding rank, while retaining more layers and attention heads (ablation) | BERT compression architecture · GLUE dev | Avg: 70.0 -> 72.6 | Helped |
| Add a second KD stage, first training a full-size student and then distilling to Architecture C (ablation) | Architecture C · GLUE dev subset | Avg: 78.6 -> 78.9 | Helped |
| Train the full-size BERT student with KD and augmented data before using it as the compressed student's teacher (ablation) | BERT_BASE teacher and Architecture C student · GLUE dev subset | Avg: 80.3 -> 81.6 | Helped |
| Iteratively prune width dimensions instead of pruning them in one step (ablation) | Architecture C with multi-stage KD · GLUE dev subset | Avg: 78.9 -> 79.0 | Helped |
| Add iterative depth pruning and a third KD stage using an eight-layer teacher (ablation) | Architecture C with iterative width pruning · GLUE dev subset | Avg: 79.0 -> 79.4 | Helped |
| Use an intermediate pruning frequency rather than pruning rapidly or slowly (ablation) | Iterative pruning during the third KD stage · CoLA and SST-2 dev | performance | Mixed |
| Use a linear-decaying rather than constant learning-rate schedule during slow pruning (ablation) | Iterative pruning during the third KD stage · CoLA dev | performance | Hurt |
| Use a linear-decaying rather than constant learning-rate schedule during slow pruning (ablation) | Iterative pruning during the third KD stage · SST-2 dev | performance | No clear effect |
| Compress BERT using ROSITA's integrated pruning, low-rank factorization, and KD framework | BERT_BASE · GLUE test average | Avg: 76.7 -> 78.0 | Helped |
| Use ROSITA rather than MobileBERT at a similar parameter budget | BERT_BASE compression · GLUE test average | Avg: 76.7 -> 78.0 | Helped |
Closest papers by meaning
- Exploring extreme parameter compression for pre-trained language models (ICLR 2022 · similarity 0.90)
- From Dense to Sparse: Contrastive Pruning for Better Pre-trained Language Model Compression (AAAI 2022 · similarity 0.87)
- DynaBERT: Dynamic BERT with Adaptive Width and Depth (NeurIPS 2020 · similarity 0.87)
- Language Modeling Is Compression (ICLR 2024 · similarity 0.86)
- Compressing LLMs: The Truth is Rarely Pure and Never Simple (ICLR 2024 · similarity 0.85)
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (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.