Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
On Episodes, Prototypical Networks, and Few-Shot 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 |
|---|---|---|---|
| replace Prototypical Network episodic loss with non-episodic NCA loss (nearest centroid eval) | Prototypical Networks · miniImageNet 1-shot | accuracy (%): 61.32 -> 62.55 | Helped |
| replace Prototypical Network episodic loss with NCA loss | Prototypical Networks · CIFAR-FS 5-shot | accuracy (%): 84.61 -> 85.15 | Helped |
| replace Prototypical Network episodic loss with NCA loss | Prototypical Networks · miniImageNet 5-shot | accuracy (%): 77.77 -> 78.27 | Mixed |
| replace Matching Network episodic loss with NCA soft assignment | Matching Networks · CIFAR-FS 1-shot | accuracy (%): 71.86 -> 72.49 | Helped |
| tune episode configuration (batch size, shots, m+n) instead of original Snell/Vinyals settings (ablation) | Prototypical Networks · miniImageNet 5-shot | accuracy (%): 75.42 -> 77.77 | Helped |
| increase the training m+n for 1-shot episodes from 8 to 32 (ablation) | Prototypical and Matching Networks · CIFAR-FS 1-shot | accuracy (%): 10% worse in the best case | Hurt |
| sample batches with replacement for NCA (ablation) | NCA · miniImageNet and CIFAR-FS val | accuracy | No clear effect |
| disable prototypes during training of PNs (5-shot) (ablation) | Prototypical Networks · miniImageNet and CIFAR-FS val | accuracy | Hurt |
| ignore support/query separation and compute loss on all pairs (ablation) | Prototypical Networks · miniImageNet and CIFAR-FS val | accuracy | Helped |
| enable all ablations (replacement sampling, no S/Q split, no prototypes) (ablation) | Prototypical Networks · miniImageNet and CIFAR-FS val | accuracy | Helped |
| randomly subsample distance pairs in NCA batch (ablation) | NCA · miniImageNet val | accuracy | Hurt |
| use nearest centroid instead of soft assignment for NCA evaluation (5-shot) (ablation) | NCA · CIFAR-FS 5-shot | accuracy (%): 83.38 -> 85.15 | Helped |
Closest papers by meaning
- Uniform Sampling over Episode Difficulty (NeurIPS 2021 · similarity 0.92)
- MELR: Meta-Learning via Modeling Episode-Level Relationships for Few-Shot Learning (ICLR 2021 · similarity 0.89)
- Meta-Learning for Semi-Supervised Few-Shot Classification (ICLR 2018 · similarity 0.89)
- Pseudo Informative Episode Construction for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.86)
- A Hierarchical Bayesian Model for Few-Shot Meta Learning (ICLR 2024 · similarity 0.85)
- Unsupervised Episode Generation for Graph Meta-learning (ICML 2024 · similarity 0.85)
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.