MetaLearningModelBase<T, TInput, TOutput>
Base Classes ยท AiDotNet.MetaLearning.Models
Abstract base class for meta-learning adapted models that wrap a base model with task-specific parameters.
Inheritance: ModelWrapperBase<T, TInput, TOutput> โ MetaLearningModelBase<T, TInput, TOutput>
Known derived types: AdaptedMetaModel<T, TInput, TOutput>, LFTAdaptedModel<T, TInput, TOutput>, MbPAAdaptedModel<T, TInput, TOutput>, NeuralProcessModel<T, TInput, TOutput>
For Beginners
Meta-learning adapted models are created by meta-learning algorithms (like MAML, ProtoNets, etc.) after adapting to a new task. They wrap a base neural network with task-specific parameters. This base class handles common concerns like serialization, gradient computation, and feature awareness by delegating to the wrapped model.
How It Works
Extends ModelWrapperBase with meta-learning-specific behavior:
training is not supported directly (use the meta-learning algorithm instead), and parameter
management is abstract so each adapted model can store its own task-specific parameters.
Methods
| Method | Summary |
|---|---|
void ApplyGradients(Vector<T> gradients, T learningRate) |
|
void Train(TInput input, TOutput expectedOutput) |