Interface IMetaLearnerOptions<T>

Namespace

AiDotNet.Interfaces

Assembly

AiDotNet.dll

Configuration options interface for meta-learning algorithms.

public interface IMetaLearnerOptions<T>

Type Parameters

T

The numeric type (e.g., float, double).

Remarks

Meta-learning algorithms use a two-loop optimization structure: - Inner loop: Fast adaptation to a specific task using support set - Outer loop: Meta-optimization to improve adaptation across all tasks

For Beginners: Think of meta-learning like learning to study effectively:

• Inner loop: How you study for a specific exam (practice problems, examples)
• Outer loop: Learning better study techniques by reflecting across many exams

The configuration controls both loops:

• InnerLearningRate: How aggressively to adapt to each task
• OuterLearningRate: How much to update the meta-parameters
• AdaptationSteps: How many gradient updates per task

Properties

AdaptationSteps

Gets the number of gradient descent steps for inner loop adaptation.

int AdaptationSteps { get; }

Property Value

int

How many times to update parameters on each task's support set. Typical values: 1 to 10. Default: 5

CheckpointFrequency

Gets the checkpoint save frequency in meta-iterations.

int CheckpointFrequency { get; }

Property Value

int

Save checkpoint every N meta-iterations. Only used if EnableCheckpointing is true. Default: 500

EnableCheckpointing

Gets whether to save checkpoints during training.

bool EnableCheckpointing { get; }

Property Value

bool

True to save model checkpoints periodically.

EvaluationFrequency

Gets the evaluation frequency in meta-iterations.

int EvaluationFrequency { get; }

Property Value

int

Run evaluation every N meta-iterations. Set to 0 to disable periodic evaluation. Default: 100

EvaluationTasks

Gets the number of evaluation tasks for periodic validation.

int EvaluationTasks { get; }

Property Value

int

How many tasks to use when evaluating model performance. Typical values: 100 to 1000. Default: 100

GradientClipThreshold

Gets the gradient clipping threshold to prevent exploding gradients.

double? GradientClipThreshold { get; }

Property Value

double?

Maximum gradient norm. Set to null or 0 to disable clipping. Typical values: 1.0 to 10.0. Default: 10.0

InnerLearningRate

Gets the inner loop learning rate for task-specific adaptation.

double InnerLearningRate { get; }

Property Value

double

The learning rate used during task adaptation on support sets. Typical values: 0.001 to 0.1. Default: 0.01

MetaBatchSize

Gets the number of tasks to sample per meta-update (meta-batch size).

int MetaBatchSize { get; }

Property Value

int

How many tasks to average over for each outer loop update. Typical values: 1 (online) to 32 (batch). Default: 4

Remarks

For Beginners: This controls how many tasks you learn from before updating your meta-parameters: - MetaBatchSize = 1: Update after every task (more noisy, faster iteration) - MetaBatchSize = 16: Update after 16 tasks (more stable, slower iteration)

NumMetaIterations

Gets the number of meta-training iterations to perform.

int NumMetaIterations { get; }

Property Value

int

How many times to perform the outer loop meta-update. Typical values: 100 to 10,000. Default: 1000

OuterLearningRate

Gets the outer loop learning rate for meta-optimization.

double OuterLearningRate { get; }

Property Value

double

The learning rate for updating meta-parameters. Typically 10x smaller than InnerLearningRate. Typical values: 0.0001 to 0.01. Default: 0.001

RandomSeed

Gets the random seed for reproducible task sampling and initialization.

int? RandomSeed { get; }

Property Value

int?

Random seed value. Set to null for non-deterministic behavior.

UseFirstOrder

Gets whether to use first-order approximation (e.g., FOMAML, Reptile).

bool UseFirstOrder { get; }

Property Value

bool

True to ignore second-order gradients, which is faster but may be less accurate. Default: false for MAML-based algorithms.

Remarks

For Beginners: First-order approximation: - True: Faster training, simpler gradients, works well in practice - False: More accurate gradients, slower, may be unstable with many inner steps

Methods

Clone()

Creates a deep copy of this options instance.

IMetaLearnerOptions<T> Clone()

Returns

IMetaLearnerOptions<T>

A new options instance with the same values.

IsValid()

Validates that the configuration is valid and sensible.

bool IsValid()

Returns

bool

True if the configuration is valid; false otherwise.