BarlowTwinsLoss<T>
Models & Types · AiDotNet.SelfSupervisedLearning.Losses
Barlow Twins Loss - redundancy reduction loss for self-supervised learning.
Inheritance: ContrastiveLossBase<T> → BarlowTwinsLoss<T>
For Beginners
Barlow Twins loss encourages the cross-correlation matrix between embeddings of two augmented views to be close to the identity matrix. This avoids both collapse and redundant features.
How It Works
Key insight: Instead of contrastive learning, Barlow Twins achieves representation quality through decorrelation - making different feature dimensions capture different information.
Loss components:
Loss formula:
where C is the cross-correlation matrix and λ is the redundancy weight.
Constructors
| Constructor | Summary |
|---|---|
BarlowTwinsLoss(double lambda, bool normalize) |
Initializes a new instance of the BarlowTwinsLoss class. |
Properties
| Property | Summary |
|---|---|
double Lambda { get; } |
Gets the lambda (redundancy reduction weight) parameter. |
Methods
| Method | Summary |
|---|---|
T OffDiagonalSum(Tensor<T> matrix) |
Computes the off-diagonal sum of a matrix (useful for monitoring). |
Tensor<T> ComputeCrossCorrelation(Tensor<T> z1, Tensor<T> z2, int batchSize) |
Computes the cross-correlation matrix between two sets of embeddings. |
Tensor<T> ComputeLoss(Tensor<T> z1, Tensor<T> z2) |
Computes the Barlow Twins loss between two views. |
ValueTuple<T, Tensor<T>, Tensor<T>> ComputeLossWithGradients(Tensor<T> z1, Tensor<T> z2) |
Computes the Barlow Twins loss with gradients for backpropagation. |