Research map / Label-efficient and robust learning / Recognition and long-tail
An Empirical Investigation of the Role of Pre-training in Lifelong Learning
Sanket Vaibhav Mehta, Darshan Patil, Sarath Chandar, Emma Strubell
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 |
|---|---|---|---|
| initialize with pre-trained weights instead of random (ablation) | ResNet-18 finetune · 5-dataset-CV | forgetting (%): 51.5 -> 38.3 | Improved |
| pre-trained init instead of random init (ablation) | DistilBERT finetune vs ER · 5-dataset-NLP | forgetting (%): 21.6 -> 16.7 | Improved |
| add SAM to ER | ER, DistilBERT-PT · 5-dataset-NLP | accuracy (%): 70.2 -> 71.1 | Improved |
| add SAM to finetune | FT, ResNet-18-R · 5-dataset-CV | forgetting (%): 51.5 -> 40.6 | Improved |
| add SAM to ER | ER, ResNet-18-R · 5-dataset-CV | accuracy (%): 50.6 -> 60.3 | Improved |
| add SAM to MC-SGD | MC-SGD, ResNet-18-R · Split CIFAR-50 | accuracy (%): 59.0 -> 59.1 | No clear effect |
| add SAM to MC-SGD | MC-SGD, ResNet-18-PT · 5-dataset-CV | accuracy (%): 81.9 -> 87.1 | Improved |
| add SAM to ER | ER, RoBERTa-base · 15-dataset-NLP | accuracy (%): 60.9 -> 62.1 | Improved |
| add SAM to ER | ER, T5-Small · 5-dataset-NLP | accuracy (%): 69.4 -> 69.9 | Improved |
| add SAM to ER on pre-trained ResNet | ER, ResNet-18-PT · Split YahooQA | accuracy (%): 89.4 -> 89.0 | No clear effect |
| pre-trained instead of random init (ablation) | finetune · 5-dataset-CV | sharpness (eps=5e-4): 2.1 -> 0.1 | Improved |
| add SAM to finetune (ablation) | FT, random init · 5-dataset-CV | sharpness (eps=5e-4): 2.1 -> 0.7 | Improved |
Closest papers by meaning
- Lifelong Learning via Progressive Distillation and Retrospection (ECCV 2018 · similarity 0.78)
- Simple Lifelong Learning Machines (TPAMI 2025 · similarity 0.74)
- Progress or Regress? Self-Improvement Reversal in Post-training (ICLR 2025 · similarity 0.73)
- Strong Teacher Not Needed? On Distillation in LLM Pretraining (NeurIPS 2026 · similarity 0.73)
- Causal-Driven Skill Prerequisite Structure Discovery (AAAI 2024 · similarity 0.72)
- To Code or Not To Code? Exploring Impact of Code in Pre-training (ICLR 2025 · similarity 0.72)
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.
