ContinualLearningStrategyType

Enums · AiDotNet.Models.Options

Specifies the continual learning strategy to use for preventing catastrophic forgetting.

For Beginners

Continual learning strategies help neural networks learn new tasks without forgetting previously learned ones. Different strategies use different approaches to balance learning new knowledge while preserving old knowledge.

Fields

Field Summary
ContinualLearningStrategyType AveragedGEM Averaged GEM - efficient variant using single averaged constraint.
ContinualLearningStrategyType EWC Elastic Weight Consolidation - penalizes changes to important weights using Fisher information.
ContinualLearningStrategyType ExperienceReplay Experience Replay - stores and replays examples from previous tasks.
ContinualLearningStrategyType GEM Gradient Episodic Memory - constrains gradients to not hurt stored examples.
ContinualLearningStrategyType GenerativeReplay Generative Replay - uses generative model to create pseudo-examples for rehearsal.
ContinualLearningStrategyType LearningWithoutForgetting Learning without Forgetting - uses knowledge distillation to preserve old predictions.
ContinualLearningStrategyType MAS Memory Aware Synapses - unsupervised importance estimation using output sensitivity.
ContinualLearningStrategyType OnlineEWC Online EWC - memory-efficient variant that maintains running Fisher estimate.
ContinualLearningStrategyType PackNet PackNet - isolates parameters per task through pruning and freezing.
ContinualLearningStrategyType ProgressiveNeuralNetworks Progressive Neural Networks - adds new columns with lateral connections for each task.
ContinualLearningStrategyType SynapticIntelligence Synaptic Intelligence - tracks weight importance online during training.
ContinualLearningStrategyType VCL Variational Continual Learning - Bayesian approach using posterior as prior.
int value__