Class SSLMetrics<T>

Namespace

AiDotNet.SelfSupervisedLearning.Evaluation

Assembly

AiDotNet.dll

Metrics for monitoring and evaluating self-supervised learning.

public class SSLMetrics<T>

Type Parameters

T

The numeric type used for computations.

Inheritance

object

SSLMetrics<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: These metrics help track the quality of SSL training and detect potential issues like representation collapse. Monitoring these during training helps ensure the model is learning useful representations.

Key metrics:

• Representation collapse: All embeddings become identical (very bad)
• Feature std: Standard deviation of features (should be positive)
• Alignment: Similarity between positive pairs (should be high)
• Uniformity: Distribution of features on hypersphere (should be uniform)

Methods

ComputeAlignment(Tensor<T>, Tensor<T>)

Computes alignment loss between positive pairs.

public T ComputeAlignment(Tensor<T> z1, Tensor<T> z2)

Parameters

z1 Tensor<T>

First view representations [batch_size, dim].

z2 Tensor<T>

Second view representations [batch_size, dim].

Returns

T

Average squared distance between positive pairs.

Remarks

Lower alignment loss means positive pairs are closer together in representation space, which is desirable.

ComputeCosineSimilarity(Tensor<T>, Tensor<T>)

Computes cosine similarity between corresponding pairs.

public T ComputeCosineSimilarity(Tensor<T> z1, Tensor<T> z2)

Parameters

z1 Tensor<T>

z2 Tensor<T>

Returns

T

ComputeEffectiveRank(Tensor<T>)

Computes the effective rank of the representation matrix.

public T ComputeEffectiveRank(Tensor<T> representations)

Parameters

representations Tensor<T>

Representations [batch_size, dim].

Returns

T

Effective rank (normalized entropy of singular values).

Remarks

Effective rank measures the "dimensionality" of the learned representations. Collapsed representations have low effective rank. Good representations should use many dimensions effectively.

ComputeFullReport(Tensor<T>, Tensor<T>)

Computes a full set of SSL metrics.

public SSLMetricReport<T> ComputeFullReport(Tensor<T> z1, Tensor<T> z2)

Parameters

z1 Tensor<T>

z2 Tensor<T>

Returns

SSLMetricReport<T>

ComputeRepresentationStd(Tensor<T>)

Computes the standard deviation of representations (collapse detection).

public T ComputeRepresentationStd(Tensor<T> representations)

Parameters

representations Tensor<T>

Representations [batch_size, dim].

Returns

T

Standard deviation per dimension, averaged.

Remarks

A low standard deviation indicates potential collapse - all representations are becoming similar. Good representations should have reasonable variance.

ComputeUniformity(Tensor<T>, double)

Computes uniformity loss (how uniformly distributed embeddings are).

public T ComputeUniformity(Tensor<T> representations, double t = 2)

Parameters

representations Tensor<T>

Representations [batch_size, dim].

t double

Temperature parameter (default: 2).

Returns

T

Uniformity metric (lower is more uniform).

Remarks

Measures how uniformly representations are distributed on the hypersphere. More uniform distributions indicate better representations that capture diverse information.

DetectCollapse(Tensor<T>, double)

Detects if representations are collapsing.

public bool DetectCollapse(Tensor<T> representations, double threshold = 0.01)

Parameters

representations Tensor<T>

Representations to check.

threshold double

Threshold for collapse detection (default: 0.01).

Returns

bool

True if collapse is detected.