Research map / Architectures and efficiency / Pruning and efficient tuning
BayesTune: Bayesian Sparse Deep Model Fine-tuning
Minyoung Kim, Timothy Hospedales
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 SGLD posterior mean instead of MAP for scale variables (ablation) | BayesTune · NLP 7-task average | average score: 76.30 -> 78.12 | Improved |
| use SGLD posterior mean instead of MAP (ablation) | BayesTune · CB | accuracy: 86.71 -> 91.25 | Improved |
| use SGLD posterior mean instead of MAP (ablation) | BayesTune · CoLA | score: 58.55 -> 60.85 | Improved |
| Bayesian posterior (BayesTune stage II) instead of two-stage L1-SP selection (ablation) | sparse fine-tuning · NLP 7-task average | average score: 72.64 -> 78.12 | Improved |
| apply sparse selection (stage II) after L1-SP training instead of using L1-SP alone (ablation) | L1-SP · NLP 7-task average | average score: 64.12 -> 72.64 | Improved |
| BayesTune vs SAM sparse fine-tuning | RoBERTa-base · NLP 7-task average | average score: 77.12 -> 78.12 | Improved |
| BayesTune vs random selection (Random) | RoBERTa-base · NLP 7-task average | average score: 74.55 -> 78.12 | Improved |
| increase sparsity level p from 0.005 to 0.5 (ablation) | BayesTune · NLP average test accuracy | accuracy | Hurt |
| vary sparsity level p among 0.01 to 0.1 (ablation) | BayesTune · NLP average test accuracy | accuracy | No clear effect |
| place scale variables on ViT backbone parameters instead of extra modules (ablation) | BayesTune · VTAB-1K | accuracy | Hurt |
| sparser adapter updates (smaller p) on Sun397, CIFAR100, DTD, Retinopathy (ablation) | BayesTune · VTAB-1K | accuracy | Improved |
| denser adapter updates on DMLab, Camelyon, SVHN, dSpr-Loc (ablation) | BayesTune · VTAB-1K | accuracy | Mixed |
Closest papers by meaning
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.89)
- Sparse is Enough in Fine-tuning Pre-trained Large Language Models (ICML 2024 · similarity 0.89)
- Be Confident in What You Know: Bayesian Parameter Efficient Fine-Tuning of Vision Foundation Models (NeurIPS 2024 · similarity 0.89)
- Sparse maximal update parameterization: A holistic approach to sparse training dynamics (NeurIPS 2024 · similarity 0.88)
- Bayesian Low-Rank Learning (Bella): A Practical Approach to Bayesian Neural Networks (AAAI 2025 · similarity 0.87)
- Training Bayesian Neural Networks with Sparse Subspace Variational Inference (ICLR 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.
