IContinualLearner<T, TInput, TOutput>

Interfaces · AiDotNet.ContinualLearning.Interfaces

Interface for continual learning trainers that can learn multiple tasks sequentially.

Known implementations: ContinualLearnerBase<T, TInput, TOutput>

For Beginners

Continual learning (also called lifelong learning) is the ability to learn new tasks over time without forgetting previously learned knowledge. Traditional neural networks suffer from “catastrophic forgetting” - when trained on new data, they forget what they learned before.

How It Works

This interface provides methods to:

  • Learn new tasks while protecting old knowledge
  • Evaluate performance on all learned tasks
  • Save and load model state for resuming training

Common Implementations:

  • EWCTrainer: Uses Elastic Weight Consolidation to protect important weights
  • LwFTrainer: Uses Learning without Forgetting with knowledge distillation
  • GEMTrainer: Uses Gradient Episodic Memory to constrain gradients

References:

  • Parisi et al. “Continual Lifelong Learning with Neural Networks: A Review” (2019)
  • De Lange et al. “A Continual Learning Survey” (2021)

Properties

Property Summary
IContinualLearnerConfig<T> Config { get; } Gets the configuration for the continual learner.
IFullModel<T, TInput, TOutput> BaseModel { get; } Gets the underlying model being trained.
bool IsTraining { get; } Gets whether the learner is currently training.
int TasksLearned { get; } Gets the number of tasks that have been learned.
long MemoryUsageBytes { get; } Gets the current memory usage of the learner in bytes.

Methods

Method Summary
ContinualEvaluationResult<T> EvaluateAllTasks() Evaluates the model on all learned tasks.
ContinualLearningResult<T> GetTaskHistory(int taskId) Gets the training history for a specific task.
ContinualLearningResult<T> LearnTask(IDataset<T, TInput, TOutput> taskData) Learns a new task from the provided data.
ContinualLearningResult<T> LearnTask(IDataset<T, TInput, TOutput> taskData, IDataset<T, TInput, TOutput> validationData, int? earlyStoppingPatience) Learns a new task with a validation set for early stopping and monitoring.
IReadOnlyDictionary<int, T> ComputeForgetting() Computes the current forgetting metric for all tasks.
IReadOnlyList<ContinualLearningResult<T>> GetAllHistory() Gets all training history.
TaskEvaluationResult<T> EvaluateTask(int taskId, IDataset<T, TInput, TOutput> testData) Evaluates the model on a specific task.
void Load(string path) Loads the learner state from a file.
void Reset() Resets the learner to its initial state.
void Save(string path) Saves the learner state to a file.