Research map / Optimization / Optimizers
Meta-Learning and Universality: Deep Representations and Gradient Descent can Approximate any Learning Algorithm
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 |
|---|---|---|---|
| Take additional test-time gradient steps from the MAML initialization | MAML · Sine-curve regression, including out-of-distribution tasks | regression performance | Helped |
| Fine-tune for 100 gradient steps instead of the 5 steps used in meta-training | MAML · Omniglot | test accuracy | No clear effect |
| Initialize randomly and train from scratch rather than starting from a MAML-learned initialization | Omniglot classifier · Omniglot, 20 examples | test accuracy | Hurt |
| Use a MAML initialization rather than a recurrent meta-learner | MAML vs. SNAIL and MetaNet · 5-way 1-shot Omniglot, tasks outside the training distribution | generalization performance | Helped |
| Use a MAML initialization rather than a recurrent meta-learner | MAML vs. SNAIL and MetaNet · 5-shot sine-curve regression, extrapolated amplitudes and phases | generalization performance | Helped |
| Increase network depth while keeping the parameter count approximately 40,000 (ablation) | MAML · Polynomial regression, degree 3 | performance | Helped |
| Increase network depth beyond one hidden layer (ablation) | Task-conditioned model · Polynomial regression, degree 3 | performance | No clear effect |
| Use a bias transformation to break symmetry in the constructed gradient-based learner (ablation) | MAML universality construction · One-shot and K-shot learning procedure approximation | representational universality | Helped |
Closest papers by meaning
- Learning To Learn by Jointly Optimizing Neural Architecture and Weights (CVPR 2022 · similarity 0.89)
- Continuous-Time Meta-Learning with Forward Mode Differentiation (ICLR 2022 · similarity 0.89)
- Gradient-Based Meta-Learning with Learned Layerwise Metric and Subspace (ICML 2018 · similarity 0.89)
- Learning Universal Predictors (ICML 2024 · similarity 0.89)
- Memory-Reduced Meta-Learning with Guaranteed Convergence (AAAI 2025 · similarity 0.88)
- First-order ANIL provably learns representations despite overparametrisation (ICLR 2024 · 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.