Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
ProtoDiff: Learning to Learn Prototypical Networks by Task-Guided Diffusion
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 diffusion model with VAE (ablation) | ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 66.63 -> 64.45 | Hurt |
| replace diffusion model with normalizing flows (ablation) | ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 66.63 -> 65.11 | Hurt |
| replace diffusion model with MLP or Transformer (non-generative) (ablation) | ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 66.63 -> 64.15 (MLP), 64.97 (Transformer) | Hurt |
| add residual prototype learning (ablation) | ProtoDiff · miniImageNet | accuracy (%) | Helped |
| increase Monte Carlo prototype samples from 1 to 50 | Meta-Baseline + ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 66.63 -> 69.14 | Helped |
| increase Monte Carlo prototype samples from 1 to 50 | FEAT + ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 68.97 -> 72.25 | Helped |
| increase samples beyond 50 (to 1000) | Meta-Baseline + ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 69.14 -> 69.11 | No clear effect |
| increase diffusion timesteps from 10 to 100 (ablation) | ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 63.98 -> 66.63 | Helped |
| increase diffusion timesteps from 100 to 1000 (ablation) | ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 66.63 -> 66.74 | No clear effect |
| increase transformer depth from 12 to 15 layers (ablation) | ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 66.63 -> 64.15 | Hurt |
| increase attention heads from 1 to 16 (ablation) | ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 63.28 -> 66.63 | Helped |
| increase DDIM interval tau from 10 to 100 (ablation) | ProtoDiff · miniImageNet 5-way 1-shot | accuracy (%): 66.63 -> 63.15 | Hurt |
Closest papers by meaning
- MetaNODE: Prototype Optimization as a Neural ODE for Few-Shot Learning (AAAI 2022 · similarity 0.88)
- Why Prototypes Collapse: Diagnosing and Preventing Partial Collapse in Prototypical Self-Supervised Learning (ICLR 2026 · similarity 0.87)
- MetaDiff: Meta-Learning with Conditional Diffusion for Few-Shot Learning (AAAI 2024 · similarity 0.86)
- Prototype Completion With Primitive Knowledge for Few-Shot Learning (CVPR 2021 · similarity 0.86)
- Learning to Propagate for Graph Meta-Learning (NeurIPS 2019 · similarity 0.85)
- Mind the Gap Between Prototypes and Images in Cross-domain Finetuning (NeurIPS 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.