Research map / Learning theory and algorithms / Prediction and decision losses
On the Statistical Benefits of Curriculum Learning
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 |
|---|---|---|---|
| Choose a source task balancing transfer distance and noise rather than always training on the target | Unstructured multitask linear regression with known transfer distances · Unstructured linear regression | minimax excess risk: Q_t^2 + dσ_t^2/N | Helped |
| Learn the curriculum adaptively when transfer distances are unknown | Unstructured multitask linear regression · Unstructured linear regression with unknown transfer distances | minimax excess risk: min{σ_T^2 log(T)/N, Q_sub^2} + min_t dσ_t^2/N | Hurt |
| Use target-task observations to identify a matching source task | Unstructured multitask linear regression with at least one identical source · Unstructured linear regression | excess risk: C_0log(T/δ)(C_2σ_T^2/N + dTσ_{t^*}^2/(C_1N)) | Helped |
| Select tasks optimistically to increase coverage of representation directions | Structured multitask linear regression · Structured linear regression | normalized diversity λ_{N,k}/N: λ/(C_4k) - √(σ^2 C_5^2 d k T log(κN/(Tδ))/(C_4^2 C_1^2 λN)) | Helped |
| Use an accurate prediction-gain scheduler with averaging SGD | Unstructured multitask linear regression with identity covariances · Unstructured linear regression | target excess risk: Δ^2_{t^*, T} + (dσ_{t^*}^2 + C_5)log(N)/N | Helped |
| Use greedy local prediction gain as a task-selection criterion | Unstructured multitask linear regression · Unstructured linear regression | one-step prediction gain | Mixed |
| Allocate samples equally across tasks in structured representation learning | Structured multitask linear regression · Structured linear regression with repeated task directions | diversity: λ_{N,k} ≈ 1/(kT) | Hurt |
| Use an oracle to select a diverse curriculum | Structured multitask linear regression · Structured linear regression | normalized diversity λ_{N,k}/N: λ/(C_4k) | Helped |
Closest papers by meaning
- When Does Curriculum Learning Help? A Theoretical Perspective (NeurIPS 2025 · similarity 0.91)
- Minimax Curriculum Learning: Machine Teaching with Desirable Difficulties and Scheduled Diversity (ICLR 2018 · similarity 0.89)
- Bandit Guided Submodular Curriculum for Adaptive Subset Selection (NeurIPS 2025 · similarity 0.86)
- CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs (ICLR 2026 · similarity 0.85)
- CurBench: Curriculum Learning Benchmark (ICML 2024 · similarity 0.85)
- Dual-Level Curriculum Meta-Learning for Noisy Few-Shot Learning Tasks (AAAI 2024 · similarity 0.84)
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.