Contrastive Divergence (CD) is an optimization method for efficiently estimating parameters in probabilistic models such as Boltzmann machines.
The method is based on approximately calculating the gradient of the log-likelihood function by starting a short Markov chain only a few steps from a data distribution, thus capturing the difference between the data distribution and the model distribution.
This significantly reduces the effort compared to exact maximum likelihood estimation.
Contrastive Divergence is often used in unsupervised learning models to discover latent structures in data and speed up training.
Especially in Restricted Boltzmann Machines (RBMs), CD is a standard method for efficiently learning the weight matrix.