FewTUREAlgorithm<T, TInput, TOutput>
Models & Types · AiDotNet.MetaLearning.Algorithms
Implementation of FewTURE (Few-shot Transformer with Uncertainty and Reliable Estimation) (Hiller et al., ECCV 2022).
Inheritance: ModelBase<T, TInput, TOutput> → MetaLearnerBase<T, TInput, TOutput> → FewTUREAlgorithm<T, TInput, TOutput>
Implements: IGeneratedParameterRegistrar<T>
For Beginners
FewTURE compares images piece by piece, not as wholes:
Standard approach: Represent each image as ONE feature vector, compare vectors. Problem: A bird’s beak might be the key difference, but it’s a tiny part of the image.
FewTURE’s approach:
- Split each image into patches (like a puzzle, e.g., 14x14 = 196 patches)
- Use a Vision Transformer to get a feature for each patch (token)
- Compare queries to support classes at the PATCH level
- Estimate uncertainty for each patch comparison
- Weight reliable patches more, uncertain patches less
Why uncertainty matters: Not all patches are equally informative:
- Background patches are mostly noise (high uncertainty)
- Discriminative patches (beak, stripes) are informative (low uncertainty)
FewTURE learns to focus on the informative patches automatically.
How It Works
FewTURE uses vision transformers with token-level local features and uncertainty estimation for few-shot classification. Instead of comparing global image features, it compares at the patch/token level and weights predictions by their estimated reliability.
Algorithm - FewTURE:
Reference: Hiller, M., Ma, R., Harber, M., & Ommer, B. (2022). Rethinking Generalization in Few-Shot Classification. ECCV 2022.
Constructors
| Constructor | Summary |
|---|---|
FewTUREAlgorithm(FewTUREOptions<T, TInput, TOutput> options) |
Initializes a new FewTURE meta-learner. |
Properties
| Property | Summary |
|---|---|
MetaLearningAlgorithmType AlgorithmType { get; } |
Methods
| Method | Summary |
|---|---|
IModel<TInput, TOutput, ModelMetadata<T>> Adapt(IMetaLearningTask<T, TInput, TOutput> task) |
|
T MetaTrain(TaskBatch<T, TInput, TOutput> taskBatch) |