ExpectedGradientLength<T, TInput, TOutput>

Models & Types · AiDotNet.ContinualLearning.Strategies

Expected Gradient Length (EGL) strategy for continual learning.

Inheritance: ContinualLearningStrategyBase<T, TInput, TOutput> → ExpectedGradientLength<T, TInput, TOutput>

For Beginners

EGL protects important parameters by measuring how much each parameter affects the output when training on a task. Parameters with larger expected gradient lengths are considered more important.

How It Works

How it works:

The Math:

EGL loss: L_total = L_task + (λ/2) * Σᵢ Ωᵢ * (θᵢ - θ*ᵢ)²

Where Ωᵢ = E[|∂L/∂θᵢ|] is the expected gradient length for parameter i

Comparison to Other Methods:

EGL is simpler than EWC and provides a direct measure of how much each parameter contributes to the loss during training.

Constructors

Constructor Summary
ExpectedGradientLength(ILossFunction<T> lossFunction, EGLOptions<T> options) Initializes a new EGL strategy with custom options.
ExpectedGradientLength(ILossFunction<T> lossFunction, T lambda, int numSamples) Initializes a new EGL strategy with default options.

Properties

Property Summary
T Lambda { get; } Gets the regularization strength (lambda).
Vector<T> Importance { get; } Gets the accumulated importance weights.
Vector<T> PreviousParameters { get; } Gets the stored parameters from previous tasks.
bool ModifiesArchitecture { get; }
bool RequiresMemoryBuffer { get; }
long MemoryUsageBytes { get; }
string Name { get; }

Methods

Method Summary
T ComputeRegularizationLoss(IFullModel<T, TInput, TOutput> model)
Vector<T> AdjustGradients(IFullModel<T, TInput, TOutput> model, Vector<T> gradients)
void FinalizeTask(IFullModel<T, TInput, TOutput> model)
void PrepareForTask(IFullModel<T, TInput, TOutput> model, IDataset<T, TInput, TOutput> taskData)
void Reset()