TaskDifficultyEstimator<T>

Helpers & Utilities · AiDotNet.MetaLearning.Data

Estimates the difficulty of a meta-learning task based on geometric properties of the data: inter-class separation, intra-class variance, and support/query alignment.

For Beginners

This utility looks at the support and query data and estimates how hard the task is. Tasks where classes overlap a lot are harder than tasks where classes are clearly separated. This estimate can feed into curriculum or dynamic samplers.

How It Works

Reference: Adaptive Task Sampling for Meta-Learning (2024).

Methods

Method Summary
T EstimateDifficulty(Vector<T> supportX, Vector<T> supportY, int numWays, int numShots) Estimates difficulty of a task in [0, 1] based on support set geometry.
T EstimateFisherDifficulty(Vector<T> supportX, Vector<T> supportY, int numWays) Estimates difficulty using the Fisher discriminant ratio: difficulty = 1 - (between-class variance / total variance).