Class DeepEnsemble<T>

Namespace

AiDotNet.UncertaintyQuantification.BayesianNeuralNetworks

Assembly

AiDotNet.dll

Implements Deep Ensembles for uncertainty estimation.

public class DeepEnsemble<T> : IUncertaintyEstimator<T>

Type Parameters

T

The numeric type used for calculations (e.g., float, double).

Inheritance

object

DeepEnsemble<T>

Implements

IUncertaintyEstimator<T>

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

For Beginners: Deep Ensembles is one of the most effective methods for uncertainty estimation.

The concept is simple: Train multiple independent neural networks (with different random initializations) on the same task, then use them all to make predictions. The diversity in their predictions gives you uncertainty estimates.

Think of it like a panel of doctors giving diagnoses:

• If all doctors agree on the diagnosis, confidence is high
• If doctors give different diagnoses, uncertainty is high

Advantages:

• Very reliable uncertainty estimates
• No special training procedures needed
• Each network can use standard architectures

Disadvantages:

• Requires training and storing multiple networks
• Slower inference (must run all networks)
• Higher memory usage

Research shows that ensembles of just 5 networks often outperform more complex Bayesian methods.

Constructors

DeepEnsemble(List<INeuralNetwork<T>>)

Initializes a new instance of the DeepEnsemble class.

public DeepEnsemble(List<INeuralNetwork<T>> models)

Parameters

models List<INeuralNetwork<T>>

The collection of trained models in the ensemble.

Remarks

For Beginners: Each model should be trained independently with different random initializations. Typically 5-10 models is a good balance between performance and cost.

Exceptions

ArgumentException

Thrown when the model list is empty.

Properties

EnsembleSize

Gets the number of models in the ensemble.

public int EnsembleSize { get; }

Property Value

int

Methods

EstimateAleatoricUncertainty(Tensor<T>)

Estimates aleatoric uncertainty from the ensemble.

public Tensor<T> EstimateAleatoricUncertainty(Tensor<T> input)

Parameters

input Tensor<T>

The input tensor.

Returns

Tensor<T>

The aleatoric uncertainty estimate.

Remarks

For Beginners: If each model in the ensemble outputs both a prediction and its own uncertainty estimate (e.g., Gaussian likelihood), the average of individual model uncertainties represents aleatoric uncertainty (data noise).

Important: This implementation does not currently have per-model variance heads, so a true aleatoric/epistemic decomposition is not available. As a conservative default, this method returns zeros; to estimate aleatoric uncertainty, modify ensemble members to output explicit variance (e.g., a Gaussian likelihood head).

EstimateEpistemicUncertainty(Tensor<T>)

Estimates epistemic uncertainty from the ensemble.

public Tensor<T> EstimateEpistemicUncertainty(Tensor<T> input)

Parameters

input Tensor<T>

The input tensor.

Returns

Tensor<T>

The epistemic uncertainty estimate.

Remarks

For Beginners: The disagreement between ensemble members represents epistemic uncertainty (model uncertainty). When models disagree, it means they're uncertain about the correct answer.

GetAllPredictions(Tensor<T>)

Gets the predictions from all ensemble members.

public List<Tensor<T>> GetAllPredictions(Tensor<T> input)

Parameters

input Tensor<T>

The input tensor.

Returns

List<Tensor<T>>

A list of predictions from each ensemble member.

Remarks

For Beginners: This is useful when you want to analyze individual model predictions or implement custom uncertainty aggregation methods.

PredictWithUncertainty(Tensor<T>)

Predicts output with uncertainty estimates from the ensemble.

public UncertaintyPredictionResult<T, Tensor<T>> PredictWithUncertainty(Tensor<T> input)

Parameters

input Tensor<T>

The input tensor.

Returns

UncertaintyPredictionResult<T, Tensor<T>>

A tuple containing the mean prediction and uncertainty.

Remarks

For Beginners: This runs the input through all models in the ensemble, then returns the average prediction and how much the models disagreed. More disagreement = higher uncertainty.