Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
TaskNorm: Rethinking Batch Normalization for Meta-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 |
|---|---|---|---|
| Blend context-set moments with instance-based moments using a learned context-size-dependent weight | CNAPs · Meta-Dataset | accuracy | Mixed |
| Use TaskNorm-I instead of the running-statistics Baseline | CNAPs · Meta-Dataset / VGG Flower | accuracy: 89.2±0.5 -> 89.6±0.6 | Helped |
| Use TaskNorm-I instead of transductive batch normalization | CNAPs · Meta-Dataset / MSCOCO | accuracy: 41.3±1.0 -> 43.4±1.0 | Helped |
| Replace local secondary moments with meta-training running moments (TaskNorm-r) (ablation) | TaskNorm · Meta-Dataset | accuracy | Hurt |
| Learn alpha shared across channels as a function of context-set size rather than as an independent shared parameter (ablation) | TaskNorm-I with CNAPs · Meta-Dataset | accuracy: 49.6±1.1 -> 50.6±1.1; 90.9±0.6 -> 90.7±0.6; 84.6±0.6 -> 83.8±0.6; 73.2±0.9 -> 74.6±0.8; 58.5±0.7 -> 62.1±0.7; 73.9±0.7 - | Helped |
| Make alpha channel-specific rather than shared across channels (ablation) | TaskNorm-I · Meta-Dataset | accuracy: 49.8±1.1 -> 50.6±1.1; 90.1±0.6 -> 90.7±0.6; 84.4±0.6 -> 83.8±0.6; 73.1±0.9 -> 74.6±0.8; 61.0±0.8 -> 62.1±0.7; 74.2±0.7 - | Hurt |
| Use channel-specific rather than shared independent alpha (ablation) | TaskNorm-I with MAML · Omniglot-5-1 | accuracy: 94.4±0.8 -> 90.7±1.0 | Hurt |
| Test all target examples together rather than one example at a time | TBN with CNAPs · Meta-Dataset / ILSVRC | accuracy: 9.5±0.3 -> 50.2±1.0 | Helped |
Closest papers by meaning
- MetaNorm: Learning to Normalize Few-Shot Batches Across Domains (ICLR 2021 · similarity 0.93)
- Meta-Learning without Memorization (ICLR 2020 · similarity 0.89)
- Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Training and Effective Adaptation (ICML 2021 · similarity 0.88)
- High-Performance Large-Scale Image Recognition Without Normalization (ICML 2021 · similarity 0.88)
- Deconstructing the Regularization of BatchNorm (ICLR 2021 · similarity 0.88)
- Prompt Learning via Meta-Regularization (CVPR 2024 · similarity 0.87)
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.