← Back to Gallery

Information Bottleneck

Information Bottleneck Method

The IB method finds compressed representation T of data X that preserves relevant information about target Y.

min I(X;T) - β·I(T;Y)

Trade-off: Low β emphasizes compression, high β preserves more task-relevant information.

Applications

• Deep learning (understanding neural networks)

• Feature extraction and dimensionality reduction

• Clustering and categorization

• Lossy compression with relevance