Table of Contents

Class BarlowTwins<T>

Namespace
AiDotNet.SelfSupervisedLearning
Assembly
AiDotNet.dll

Barlow Twins: Self-Supervised Learning via Redundancy Reduction.

public class BarlowTwins<T> : SSLMethodBase<T>, ISSLMethod<T>

Type Parameters

T

The numeric type used for computations.

Inheritance
BarlowTwins<T>
Implements
Inherited Members

Remarks

For Beginners: Barlow Twins learns representations by making the cross-correlation matrix between embeddings of two augmented views close to the identity matrix. This achieves both invariance (diagonal = 1) and reduces redundancy between features (off-diagonal = 0).

Key innovations:

  • No negatives: Unlike contrastive methods, doesn't need negative samples
  • No momentum encoder: Uses a simpler symmetric architecture
  • Decorrelation: Explicitly reduces redundancy between feature dimensions
  • No asymmetry: Both branches are identical (no predictor needed)

Loss components:

  • Invariance term: Makes diagonal elements = 1 (same view features match)
  • Redundancy reduction: Makes off-diagonal elements = 0 (decorrelation)

Cross-correlation matrix:

C_ij = (1/N) * Σ_b z1[b,i] * z2[b,j]
Loss = Σ_i (1 - C_ii)² + λ * Σ_i Σ_{j≠i} C_ij²

Reference: Zbontar et al., "Barlow Twins: Self-Supervised Learning via Redundancy Reduction" (ICML 2021)

Constructors

BarlowTwins(INeuralNetwork<T>, IProjectorHead<T>, SSLConfig?)

Initializes a new instance of the BarlowTwins class.

public BarlowTwins(INeuralNetwork<T> encoder, IProjectorHead<T> projector, SSLConfig? config = null)

Parameters

encoder INeuralNetwork<T>

The encoder network (shared between both branches).

projector IProjectorHead<T>

Projection head (typically 3-layer MLP with 8192 output dim).

config SSLConfig

Optional SSL configuration.

Properties

Category

Gets the category of this SSL method.

public override SSLMethodCategory Category { get; }

Property Value

SSLMethodCategory

Remarks

Categories include Contrastive, NonContrastive, Generative, and SelfDistillation.

Lambda

Gets the lambda (redundancy reduction weight) used in the loss.

public double Lambda { get; }

Property Value

double

Name

Gets the name of this SSL method.

public override string Name { get; }

Property Value

string

Remarks

Examples: "SimCLR", "MoCo v2", "BYOL", "DINO", "MAE"

RequiresMemoryBank

Indicates whether this method requires a memory bank for negative samples.

public override bool RequiresMemoryBank { get; }

Property Value

bool

Remarks

For Beginners: Memory banks store embeddings from previous batches to use as negative samples in contrastive learning. MoCo uses this, SimCLR does not.

UsesMomentumEncoder

Indicates whether this method uses a momentum-updated encoder.

public override bool UsesMomentumEncoder { get; }

Property Value

bool

Remarks

For Beginners: A momentum encoder is a slowly-updated copy of the main encoder. Methods like MoCo, BYOL, and DINO use this to provide stable targets.

Methods

Create(INeuralNetwork<T>, int, int, int, double)

Creates a Barlow Twins instance with default configuration.

public static BarlowTwins<T> Create(INeuralNetwork<T> encoder, int encoderOutputDim, int projectionDim = 8192, int hiddenDim = 8192, double lambda = 0.0051)

Parameters

encoder INeuralNetwork<T>

The backbone encoder.

encoderOutputDim int

Output dimension of the encoder.

projectionDim int

Dimension of the projection space (default: 8192).

hiddenDim int

Hidden dimension of the projector MLP (default: 8192).

lambda double

Redundancy reduction weight (default: 0.0051).

Returns

BarlowTwins<T>

A configured Barlow Twins instance.

TrainStepCore(Tensor<T>, SSLAugmentationContext<T>?)

Implementation-specific training step logic.

protected override SSLStepResult<T> TrainStepCore(Tensor<T> batch, SSLAugmentationContext<T>? augmentationContext)

Parameters

batch Tensor<T>

The input batch tensor.

augmentationContext SSLAugmentationContext<T>

Optional augmentation context.

Returns

SSLStepResult<T>

The result of the training step.

Edit this page