Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Zero-Shot Task Adaptation with Relevant Feature Information
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Meta-learns a neural classifier that, for unseen target tasks with only a few relevant input features, generates task-specific linear weights by responding to feature perturbations on synthetic data, optionally adding unlabeled target data.
Problem
Learning classifiers for unseen target tasks when no labeled or unlabeled target data exist, only a few relevant input features per class are known.
Approach
Meta-learn an embedding network whose last linear layer is adapted, in the inner loop, so that predictions on synthetic data respond correctly (increase/decrease) to perturbations of the given relevant features; this lets the model reflect feature discriminability without any real target data.
Key results
- Without target unlabeled data, the proposed method averaged 67.89% test accuracy across datasets and J settings, best or comparable in 7 of 8 cases (vs. 64.66% for Metric and 64.26% for Proto).
- With unlabeled target data, OursU averaged 70.73%, outperforming LRU (67.08%), LRD2U (68.68%), ProtoU (66.46%) and NPU (57.44%).
- Ablation: removing meta-learning (Target) dropped average accuracy to 58.73%, showing the value of meta-learning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use meta-learning from source tasks rather than training on target relevant features alone (ablation) | Ours · 20News, WoS, URL, and Mnistr (J=2,4), no unlabeled target data | accuracy: 62.46 vs 53.05; 70.73 vs 52.11; 53.29 vs 50.47; 77.23 vs 73.47 (J=2); 65.17 vs 53.72; 77.49 vs 54.62; 53.48 vs 50.61; 83 | Helped |
| Adapt to labeled data in the inner optimization instead of simulating test-time relevant-feature adaptation (ablation) | Ours · 20News and WoS (J=2), no unlabeled target data | accuracy: 62.46 vs 50.01; 70.73 vs 51.69 | Hurt |
| Use real unlabeled data instead of synthetic data in the inner optimization (ablation) | Ours · 20News and WoS (J=2), no unlabeled target data | accuracy: 62.46 vs 61.01; 70.73 vs 69.52 | Hurt |
| Increase the number of synthetic data (ablation) | Ours · 20News, WoS, URL, and Mnistr | accuracy | Helped |
| Add unlabeled target data to the adaptation procedure | Ours · 20News, WoS, URL, and Mnistr (J=2,4) | accuracy: 62.46 -> 65.03; 65.17 -> 69.28; 70.73 -> 73.31; 77.49 -> 83.92; 53.29 -> 52.20; 53.48 -> 51.02; 77.23 -> 83.10; 83.25 -> | Mixed |
| Add the proposed relevant-feature regularizer to unlabeled-data adaptation | LRD2U / OursU · 20News, WoS, URL, and Mnistr (J=2,4) | accuracy: 61.33 -> 65.03; 65.09 -> 69.28; 76.88 -> 73.31; 76.67 -> 83.92; 49.63 -> 52.20; 49.65 -> 51.02; 82.23 -> 83.10; 87.77 -> | Mixed |
Surprising findings
- Keyword matching and kNN using only relevant features performed poorly on most datasets, indicating relevant features alone are generally insufficient for classification.
- On Mnist-r with J=4, the proposed method performed worse than kNN (83.25 vs 85.92).
Limitations
- Assumes a simple monotonic relationship between relevant features and class labels; more complex prior knowledge is not handled.
- Performance is weak on some datasets (e.g., URL near 53%) and on Mnist-r with J=4 transfer is weaker than kNN.
- Requires relevant features for source tasks, which may need to be estimated via l1-regularized logistic regression.
Experimental setup
Datasets: 20News WoS URL Mnist-r
Benchmarks: 20News WoS URL Mnist-r
Metrics: average test accuracy
Baselines: KW kNN LR NN LRD2 Proto NP Relation Metric LRU LRD2U ProtoU
Builds on: MAML R2-D2 (differentiable ridge/logistic regression meta-learning) prototypical networks dataless classification
Closest papers by meaning
- Meta-Learning without Memorization (ICLR 2020 · similarity 0.90)
- Zero-Shot Task Transfer (CVPR 2019 · similarity 0.89)
- Task Aligned Generative Meta-learning for Zero-shot Learning (AAAI 2021 · similarity 0.88)
- Task Agnostic Meta-Learning for Few-Shot Learning (CVPR 2019 · similarity 0.88)
- A Generalization Theory for Zero-Shot Prediction (ICML 2025 · similarity 0.87)
- Task-Specific Gradient Adaptation for Few-Shot One-Class Classification (CVPR 2025 · 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.