Maximally Collapsing Metric Learning (MCML)

An information-theory based distance metric learning algorithm. It obtains a metric that minimizes the Kullback-Leibler divergence to the ideal distribution where every points in the same class collapse into a single point, and different class points are infinitely far away.

Watch the full MCML documentation here.

References

Amir Globerson and Sam T Roweis. “Metric learning by collapsing classes”. In: Advances in neural information processing systems. 2006, pages 451-458.