FedSamAggregationStrategy<T>

Models & Types · AiDotNet.FederatedLearning.Aggregators

Implements FedSAM (Sharpness-Aware Minimization for Federated Learning) aggregation strategy.

Inheritance: FederatedLearningComponentBase<T> → AggregationStrategyBase<Dictionary<string, T[]>, T> → ParameterDictionaryAggregationStrategyBase<T> → FedSamAggregationStrategy<T>

For Beginners

Regular optimization finds a low point (minimum) in the loss landscape, but this point might be “sharp” — a small change in parameters causes a large change in loss. FedSAM instead seeks “flat” minima that are more robust, which is especially important in FL where each client’s data creates a different loss landscape.

How It Works

Local training uses a two-step process per batch:

Variants:

Reference: Caldarola, D., et al. (2022). “Improving Generalization in Federated Learning by Seeking Flat Minima.”

Constructors

Constructor Summary
FedSamAggregationStrategy(double perturbationRadius, FedSamVariant variant, double globalPerturbationRadius, double approximationCoeff, double controlCoeff) Initializes a new instance of the FedSamAggregationStrategy class.

Properties

Property Summary
FedSamVariant Variant { get; } Gets the FedSAM variant being used.
double ApproximationCoefficient { get; } Gets the approximation coefficient (FedSpeed/FedLESAM).
double ControlCoefficient { get; } Gets the stochastic control coefficient (FedSCAM).
double GlobalPerturbationRadius { get; } Gets the global perturbation radius (FedSMOO).
double PerturbationRadius { get; } Gets the SAM perturbation radius (rho).

Methods

Method Summary
Dictionary<string, T[]> Aggregate(Dictionary<int, Dictionary<string, T[]>> clientModels, Dictionary<int, double> clientWeights) Aggregates client models using standard weighted averaging.
Dictionary<string, T[]> ApplyPerturbation(Dictionary<string, T[]> parameters, Dictionary<string, T[]> perturbation) Applies a perturbation to model parameters: w_perturbed = w + epsilon.
Dictionary<string, T[]> ComputePerturbation(Dictionary<string, T[]> gradient) Computes the SAM perturbation direction from the current gradient.
Dictionary<string, T[]> ComputeSAMGradient(Dictionary<string, T[]> currentGradient, Dictionary<string, T[]> perturbedGradient) Computes the complete SAM-modified gradient for a training step.
string GetStrategyName()
void UpdateGradientHistory(Dictionary<string, T[]> gradient) Updates the gradient history (used by FedSpeed/FedLESAM for approximation).