CAMLAlgorithm<T, TInput, TOutput>
Models & Types · AiDotNet.MetaLearning.Algorithms
Implementation of CAML (Context-Aware Meta-Learning) (Fifty et al., NeurIPS 2023).
Inheritance: ModelBase<T, TInput, TOutput> → MetaLearnerBase<T, TInput, TOutput> → CAMLAlgorithm<T, TInput, TOutput>
Implements: IGeneratedParameterRegistrar<T>
For Beginners
CAML is built on a simple but powerful insight:
The insight: Modern pretrained models (CLIP, DINO) produce such good features that you don’t need to fine-tune them. Instead, learn a lightweight context module that adapts how you USE the features for each specific task.
How it works:
- Extract features using a frozen pretrained backbone (no gradient computation needed)
- Compute class prototypes from support features (like ProtoNets)
- Pass prototypes through a context module that adjusts them based on the task
- Classify queries by distance to context-adapted prototypes
Why freeze the backbone?
- Much faster training (no backbone gradients)
- Avoids overfitting on small support sets
- Preserves the rich representations learned during pretraining
- Only the small context module needs to be meta-learned
How It Works
CAML uses a frozen pretrained backbone with a lightweight context module that adapts features based on the support set. Classification is performed by comparing query features to context-adapted class prototypes.
Algorithm - CAML:
Reference: Fifty, C., Duan, D., Junkins, R.G., Amid, E., Leskovec, J., Re, C., & Thrun, S. (2023). Context-Aware Meta-Learning. NeurIPS 2023.
Constructors
| Constructor | Summary |
|---|---|
CAMLAlgorithm(CAMLOptions<T, TInput, TOutput> options) |
Initializes a new CAML 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) |