BADGEStrategy<T, TInput, TOutput>
Models & Types · AiDotNet.ActiveLearning.Strategies.Hybrid
BADGE (Batch Active learning by Diverse Gradient Embeddings) strategy.
Implements: IQueryStrategy<T, TInput, TOutput>
For Beginners
BADGE is a state-of-the-art strategy that combines uncertainty and diversity. It uses gradient embeddings (gradients of the loss with respect to the final layer) to represent samples, then uses k-means++ to select diverse samples.
How It Works
How BADGE Works:
Key Insight: Gradient embeddings naturally combine model uncertainty (large gradients for uncertain samples) with feature diversity (different gradient directions for different samples).
When to Use:
Reference: Ash et al. “Deep Batch Active Learning by Diverse, Uncertain Gradient Lower Bounds” (ICLR 2020)
Constructors
| Constructor | Summary |
|---|---|
BADGEStrategy(int numClasses) |
Initializes a new BADGE strategy. |
BADGEStrategy(int numClasses, ActiveLearnerConfig<T> config) |
Initializes a new BADGE strategy with configuration. |
Properties
| Property | Summary |
|---|---|
string Description { get; } |
|
string Name { get; } |
Methods
| Method | Summary |
|---|---|
Vector<T> ComputeScores(IFullModel<T, TInput, TOutput> model, IDataset<T, TInput, TOutput> unlabeledPool) |
|
int[] SelectSamples(IFullModel<T, TInput, TOutput> model, IDataset<T, TInput, TOutput> unlabeledPool, int batchSize) |
|
void Reset() |
|
void UpdateState(int[] newlyLabeledIndices, TOutput[] labels) |