iTAMLAlgorithm<T, TInput, TOutput>

Models & Types · AiDotNet.MetaLearning.Algorithms

Implementation of iTAML: incremental Task-Agnostic Meta-Learning (Rajasegaran et al., CVPR 2020).

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

How It Works

iTAML keeps a single set of meta-parameters balanced across all tasks seen so far. Each task in the batch takes r inner steps from the shared parameters; the meta-update then moves the shared parameters towards the average of the task-adapted ones, by a weight eta that shrinks as training progresses, so later tasks cannot pull the parameters away from earlier ones as fast as the first did.

Algorithm (paper Algorithm 1):

t / T is the fraction of NumMetaIterations completed. Inference-time adaptation (Algorithm 3) is the same r-step update on the task’s support set.

Constructors

Constructor Summary
iTAMLAlgorithm(iTAMLOptions<T, TInput, TOutput> options)

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)