Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Task-Specific Gradient Adaptation for Few-Shot One-Class Classification
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.
Proposes TSGA, a meta-learning method for few-shot one-class classification that learns a loss network and a task-specific adapter network to generate stable gradients and per-task learning rates/weight decay during adaptation.
Problem
Optimization-based meta-learning for few-shot one-class classification suffers from misalignment between cross-entropy loss and OCC objectives, causing unstable gradients, and from task-agnostic adaptation that limits generalization to novel tasks.
Approach
A learnable instance-wise loss network (taking predictions and representations of one-class support samples) replaces cross-entropy in the inner loop, and a Transformer-based adapter network generates task-specific per-layer learning rates and weight decay, extending OC-MAML.
Key results
- TSGA with Conv-4 reaches 73.41% on miniImageNet 2-shot and 77.73% in 10-shot, improving OC-MAML by 4.31 and 1.53 points respectively.
- TSGA improves 10-shot CIFAR-FS over OC-MAML by 2.60 points (81.70 vs 79.1) and outperforms AHDD by 1.48 and 1.50 points on miniImageNet and CIFAR-FS 2-shot.
- In one-shot OCC, TSGA reaches 71.67% on miniImageNet versus 68.89% for AHDD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add adapter network A only (ablation) | OC-MAML · miniImageNet, Omniglot, and CIFAR-FS, 2-shot | accuracy: 69.1 -> 72.13; 96.6 -> 97.89; 70.0 -> 75.91 | Helped |
| add adapter network A only (ablation) | OC-MAML · miniImageNet, Omniglot, and CIFAR-FS, 10-shot | accuracy: 76.2 -> 76.46; 97.6 -> 98.31; 79.1 -> 81.11 | Helped |
| add loss network N only (ablation) | OC-MAML · miniImageNet, Omniglot, and CIFAR-FS, 2-shot | accuracy: 69.1 -> 72.09; 96.6 -> 97.99; 70.0 -> 75.59 | Helped |
| add loss network N only (ablation) | OC-MAML · miniImageNet, Omniglot, and CIFAR-FS, 10-shot | accuracy: 76.2 -> 76.66; 97.6 -> 98.45; 79.1 -> 81.13 | Helped |
| add both adapter network A and loss network N (ablation) | OC-MAML · miniImageNet, Omniglot, and CIFAR-FS, 2-shot and 10-shot | accuracy: (+4.31); (+1.51); (+6.84); (+1.53); (+0.94); (+2.60) | Helped |
| remove feature representations from loss-network input (ablation) | TSGA loss network · miniImageNet, Omniglot, and CIFAR-FS, 2-shot and 10-shot | accuracy: 73.41 ± 0.4; 98.11 ± 0.1; 76.84 ± 0.4; 77.73 ± 0.4; 98.54 ± 0.1; 81.70 ± 0.4 | Mixed |
| use label predictions without feature representations (ablation) | TSGA loss network · miniImageNet, Omniglot, and CIFAR-FS, 2-shot and 10-shot | accuracy: 71.13 ± 0.4; 97.95 ± 0.1; 76.66 ± 0.4; 77.60 ± 0.4; 98.27 ± 0.1; 81.11 ± 0.4 | Mixed |
| use both feature representations and label predictions in the loss network (ablation) | TSGA loss network · miniImageNet, Omniglot, and CIFAR-FS, 2-shot and 10-shot | accuracy: 72.15 ± 0.4; 97.91 ± 0.1; 77.11 ± 0.4; 77.19 ± 0.4; 98.60 ± 0.1; 80.66 ± 0.4 | Mixed |
| add cross-entropy loss alongside the loss network (ablation) | TSGA loss network · miniImageNet, Omniglot, and CIFAR-FS, 2-shot and 10-shot | accuracy: 72.15 ± 0.4; 97.91 ± 0.1; 77.11 ± 0.4; 77.19 ± 0.4; 98.60 ± 0.1; 80.66 ± 0.4 | No clear effect |
| vary adapter network depth and hidden dimension (ablation) | TSGA adapter network · CIFAR-FS, 10-shot | accuracy: 81.60; 81.84; 81.39; 81.99; 81.28; 80.44; 81.47; 80.65 | No clear effect |
Surprising findings
- Adding the loss network without the feature representation input reduces gains, while adding cross-entropy on top of the loss network gives no significant change.
Limitations
- Gains on some settings are modest (e.g., 0.47 vs 0.42 cosine similarity on miniImageNet, and ResNet-12 OC-MAML reproduction is close on some settings).
- Results are reported from the authors' own reproduction for some baselines; the CNC dataset evaluation uses only three test operations.
Experimental setup
Datasets: Omniglot miniImageNet CIFAR-FS CNC Milling Machine dataset
Benchmarks: miniImageNet Omniglot CIFAR-FS CNC Milling Machine dataset
Metrics: accuracy (%) cosine similarity of support and query gradients
Baselines: OC-ProtoNet OW-ProtoNet AHDD OC-MetaOptNet OC-MetaSGD OC-MAML MAML
Builds on: MAML OC-MAML Meta-SGD ALFA AHDD
Closest papers by meaning
- Few-Shot One-Class Classification via Meta-Learning (AAAI 2021 · similarity 0.89)
- Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Training and Effective Adaptation (ICML 2021 · similarity 0.87)
- Task Agnostic Meta-Learning for Few-Shot Learning (CVPR 2019 · similarity 0.86)
- Zero-Shot Task Adaptation with Relevant Feature Information (AAAI 2024 · similarity 0.85)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.85)
- Task-Specific Preconditioner for Cross-Domain Few-Shot Learning (AAAI 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.