OnlineEWC<T>
Models & Types · AiDotNet.ContinualLearning
Implements Online Elastic Weight Consolidation (Online EWC) for continual learning.
Implements: IContinualLearningStrategy<T>
For Beginners
Online EWC is a memory-efficient variant of EWC that maintains a single running approximation of the Fisher Information Matrix across all tasks, rather than storing separate matrices for each task.
How It Works
How it works:
Formula:
F̃ = γ * F_old + F_new
θ̃* = (γ * F_old * θ*_old + F_new * θ_new) / (γ * F_old + F_new)
Advantages:
Reference: Schwarz, J. et al. “Progress & Compress: A scalable framework for continual learning” (2018). ICML.
Constructors
| Constructor | Summary |
|---|---|
OnlineEWC(double lambda, double gamma) |
Initializes a new instance of the OnlineEWC class. |
Properties
| Property | Summary |
|---|---|
bool AccumulatesAcrossTasks { get; } |
|
double Gamma { get; } |
Gets the decay factor for old Fisher information. |
double Lambda { get; set; } |
|
int TaskCount { get; } |
Gets the number of tasks processed. |
Methods
| Method | Summary |
|---|---|
T ComputeLoss(INeuralNetwork<T> network) |
|
Vector<T> ModifyGradients(INeuralNetwork<T> network, Vector<T> gradients) |
|
void AfterTask(INeuralNetwork<T> network, ValueTuple<Tensor<T>, Tensor<T>> taskData, int taskId) |
|
void BeforeTask(INeuralNetwork<T> network, int taskId) |
|
void Reset() |