PTMAPAlgorithm<T, TInput, TOutput>

Models & Types · AiDotNet.MetaLearning.Algorithms

Implementation of PT+MAP (Power Transform + Maximum A Posteriori) (Hu et al., ICLR 2021).

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

For Beginners

PT+MAP is elegantly simple:

Step 1 - Power Transform: Raw features from neural networks often have skewed distributions. The power transform x_new = sign(x) * |x|^beta makes them more bell-shaped (Gaussian). With beta=0.5, this is essentially a square root transform.

Step 2 - Center and Normalize: After the transform, center features (subtract mean) and L2 normalize.

Step 3 - MAP Estimation (transductive): Given the Gaussian assumption, compute the optimal (MAP) class assignments for ALL query examples simultaneously. This iterates between:

  • Assign queries to nearest class (E-step)
  • Update class means using assigned queries (M-step)

Why it works so well: The power transform fixes the main problem: features aren’t Gaussian. Once they ARE Gaussian, the simple MAP classifier is provably optimal. Sometimes the simplest math wins.

How It Works

PT+MAP applies a power transform to normalize feature distributions, then uses MAP estimation for transductive few-shot classification. The power transform makes features more Gaussian, enabling a simple Bayesian classifier to work well.

Algorithm - PT+MAP:

Reference: Hu, Y., Gripon, V., & Pateux, S. (2021). Leveraging the Feature Distribution in Transfer-based Few-Shot Learning. ICLR 2021.

Constructors

Constructor Summary
PTMAPAlgorithm(PTMAPOptions<T, TInput, TOutput> options) Initializes a new PT+MAP 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)