Research map / Optimization / Implicit bias and generalization
Curriculum Learning by Transfer Learning: Theory and Experiments with Deep Networks
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 |
|---|---|---|---|
| Prioritize easier examples at the beginning of training, then gradually include harder examples | Transfer curriculum learning · Deep-network classification tasks | training convergence speed | Helped |
| Use curriculum learning instead of ordinary training | Large CNN · CIFAR-100 task 1 | final test accuracy | No clear effect |
| Use curriculum learning instead of ordinary training | Small CNN · CIFAR-100 task 1 | final test accuracy | Helped |
| Use curriculum learning instead of ordinary training | CNN · STL-10 | final test accuracy | Helped |
| Use curriculum learning instead of ordinary training | CNN · CIFAR-100 tasks with differing inherent difficulty | final test accuracy | Helped |
| Use curriculum learning under excessive regularization | CNN · CIFAR-100 task 1 | performance robustness | Helped |
| Rank examples randomly rather than by estimated difficulty (ablation) | Curriculum learning · Deep-network classification tasks | training convergence speed | Hurt |
| Present more difficult examples first (ablation) | Curriculum learning · Deep-network classification tasks | training convergence speed | Hurt |
| Use a smaller classification network (ablation) | CNN with curriculum learning · CIFAR-100 task 1 | final generalization performance | Mixed |
| Sample only examples with lower ideal difficulty scores | SGD for linear regression · Linear regression | convergence rate | Helped |
| Among examples with fixed difficulty, select examples with higher loss under the current hypothesis | SGD for linear regression · Linear regression with fixed example difficulty | convergence rate | Helped |
| Prefer examples with higher current loss without fixing their difficulty | SGD for linear regression · Linear regression with continuous data distribution and realizable optimum | convergence rate | Hurt |
Closest papers by meaning
- On The Power of Curriculum Learning in Training Deep Networks (ICML 2019 · similarity 0.93)
- When Does Curriculum Learning Help? A Theoretical Perspective (NeurIPS 2025 · similarity 0.92)
- Provable Advantage of Curriculum Learning on Parity Targets with Mixed Inputs (NeurIPS 2023 · similarity 0.88)
- Automated Curriculum Learning for Neural Networks (ICML 2017 · similarity 0.88)
- Minimax Curriculum Learning: Machine Teaching with Desirable Difficulties and Scheduled Diversity (ICLR 2018 · similarity 0.88)
- An Analytical Theory of Curriculum Learning in Teacher-Student Networks (NeurIPS 2022 · similarity 0.88)
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.