ConstellationNetAlgorithm<T, TInput, TOutput>
Models & Types · AiDotNet.MetaLearning.Algorithms
Implementation of ConstellationNet (structured part-based few-shot learning) (Xu et al., ICLR 2021).
Inheritance: ModelBase<T, TInput, TOutput> → MetaLearnerBase<T, TInput, TOutput> → ConstellationNetAlgorithm<T, TInput, TOutput>
Implements: IGeneratedParameterRegistrar<T>
For Beginners
ConstellationNet recognizes objects by their PARTS and ARRANGEMENT:
The insight: A face isn’t just “eyes + nose + mouth” - it’s the specific ARRANGEMENT of these parts. Similarly, a bird isn’t just “beak + wing + tail” - it’s how they’re positioned. ConstellationNet captures both the parts AND their spatial arrangement.
How it works:
- Part detection: For each example, detect K discriminative parts
- Each part has a feature vector (what it looks like)
- Each part has a position (where it is)
- Constellation formation: Model relationships between parts
- Pairwise spatial relationships between all parts
- Creates a “constellation” = graph of parts + spatial edges
- Constellation matching: Compare query and support constellations
- Match parts between query and support (feature similarity)
- Compare spatial arrangements (structural similarity)
- Combined score determines classification
Why constellations help: Two classes of birds might share similar colors but differ in proportions. A constellation captures “beak-to-eye distance” and “wing-to-tail ratio” naturally.
How It Works
ConstellationNet detects discriminative parts in examples and models their spatial relationships as “constellations.” Classification is performed by matching the constellation structure between queries and support examples.
Algorithm - ConstellationNet:
Constructors
| Constructor | Summary |
|---|---|
ConstellationNetAlgorithm(ConstellationNetOptions<T, TInput, TOutput> options) |
Initializes a new ConstellationNet 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) |