OpenMAMLAlgorithm<T, TInput, TOutput>
Models & Types · AiDotNet.MetaLearning.Algorithms
Implementation of Open-MAML (open-set MAML with out-of-distribution detection).
Inheritance: ModelBase<T, TInput, TOutput> → MetaLearnerBase<T, TInput, TOutput> → OpenMAMLAlgorithm<T, TInput, TOutput>
For Beginners
Standard few-shot learning assumes all query examples belong to one of the support classes. But in the real world, you might encounter unknown examples.
The problem: You’re given 5 classes of animals to learn. Then a query comes in - it might be one of those 5 animals, or it might be something completely different (like a car). Standard few-shot methods will force a classification into one of the 5 classes.
How Open-MAML fixes this:
- Train with MAML as usual for inner-loop adaptation
- During meta-training, some tasks include OOD examples (open-set fraction)
- Learn a confidence threshold: if max prediction probability < threshold, reject
- The model learns to be uncertain about OOD examples while confident about in-distribution ones
Key difference from MAML: Open-MAML adds OOD detection training, so the adapted model can say “I don’t know” instead of being forced to choose a known class.
How It Works
Open-MAML extends MAML to handle open-set scenarios where query examples may belong to classes not present in the support set. It adds a confidence-based rejection mechanism to detect out-of-distribution (OOD) examples.
Algorithm - Open-MAML:
Constructors
| Constructor | Summary |
|---|---|
OpenMAMLAlgorithm(OpenMAMLOptions<T, TInput, TOutput> options) |
Initializes a new Open-MAML 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) |