FEATAlgorithm<T, TInput, TOutput>

Models & Types · AiDotNet.MetaLearning.Algorithms

Implementation of FEAT (Few-shot Embedding Adaptation with Transformer).

Inheritance: ModelBase<T, TInput, TOutput> → MetaLearnerBase<T, TInput, TOutput> → FEATAlgorithm<T, TInput, TOutput>

Implements: IGeneratedParameterRegistrar<T>

For Beginners

FEAT makes prototypes smarter by letting them “see” each other:

The insight: ProtoNets computes each class prototype in isolation. But if you know that class A and class B are very similar, you should push their prototypes apart to avoid confusion. FEAT uses a transformer to automatically learn these adjustments.

How it works:

  1. Compute initial prototypes (mean of support features per class, like ProtoNets)
  2. Feed ALL prototypes through a transformer
  • The transformer uses self-attention so each prototype can “see” all others
  • It learns to adjust prototypes based on the specific set of classes
  1. Use the adapted prototypes for nearest-prototype classification
  2. Train with both classification loss AND contrastive loss

Why the transformer helps:

  • In a 5-way task with dogs vs cats: Prototypes need to capture species differences
  • In a 5-way task with dog breeds: Same dog features need to capture breed differences
  • The transformer adjusts prototypes based on what’s needed for THIS specific task

How It Works

FEAT uses a set-to-set transformer to adapt class prototypes based on inter-class relationships. Initial prototypes (class means) are fed through the transformer, which outputs task-adapted prototypes that are more discriminative.

Algorithm - FEAT:

Reference: Ye, H.J., Hu, H., Zhan, D.C., & Sha, F. (2020). Few-Shot Learning via Embedding Adaptation with Set-to-Set Functions. CVPR 2020.

Constructors

Constructor Summary
FEATAlgorithm(FEATOptions<T, TInput, TOutput> options) Initializes a new FEAT meta-learner.

Properties

Property Summary
MetaLearningAlgorithmType AlgorithmType { get; }

Methods

Method Summary
IModel<TInput, TOutput, ModelMetadata<T>> Adapt(IMetaLearningTask<T, TInput, TOutput> task) Adapts to a new task using transformer-adapted prototypes.
T MetaTrain(TaskBatch<T, TInput, TOutput> taskBatch) Performs one meta-training step for FEAT.