SetFeatAlgorithm<T, TInput, TOutput>
Models & Types · AiDotNet.MetaLearning.Algorithms
Implementation of SetFeat (set-feature based few-shot learning) (Afrasiyabi et al., CVPR 2022).
Inheritance: ModelBase<T, TInput, TOutput> → MetaLearnerBase<T, TInput, TOutput> → SetFeatAlgorithm<T, TInput, TOutput>
Implements: IGeneratedParameterRegistrar<T>
For Beginners
SetFeat treats each class as a SET, not just a single point:
The problem with prototypes: ProtoNets computes the MEAN of support examples. This throws away information about HOW the class varies. Two classes might have the same mean but very different spreads.
How SetFeat fixes this:
- Extract features for each class’s support examples
- Feed ALL examples (as a set) into a set encoder
- The set encoder captures rich information: mean, variance, relationships
- Optional cross-attention lets classes “see” each other for context
- The resulting set-features are used for classification
Example: If you have 5 examples of cats (tabby, persian, siamese, calico, sphinx):
- ProtoNets: Average them into one “generic cat” point
- SetFeat: Encodes that cats come in different fur patterns and body types
This extra information helps distinguish cats from similar classes like small dogs.
How It Works
SetFeat learns set-level features by processing each class’s support examples as a set rather than individual instances. A set encoder with optional cross-attention computes class representations that capture intra-class variation.
Algorithm - SetFeat:
Reference: Afrasiyabi, A., Larochelle, H., Lalonde, J.F., & Gagne, C. (2022). Matching Feature Sets for Few-Shot Image Classification. CVPR 2022.
Constructors
| Constructor | Summary |
|---|---|
SetFeatAlgorithm(SetFeatOptions<T, TInput, TOutput> options) |
Initializes a new SetFeat 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) |