Beginning Anomaly Detection Using

unsupervised learning algorithm.

As for the training algorithm, there are two choices: 

contrastive divergence (CD) 

and 

persistent contrastive divergence (PCD). These algorithms both use Markov 

chains to help the training algorithm determine what direction to perform the gradient 

calculations in, but both differ and have their pros and cons. PCD can get better 

samples of the data and explore the domain of the input space better, but CD is better at 

extracting features.

Chapter 5   Boltzmann maChines

---

187

Some RBMs might also incorporate a feature known as 

momentum, which basically 

allows for an increase in learning speed and can be thought of as simulating a ball rolling 

down a hill in terms of optimizing the target function. (Think back to gradient descent 

and how the goal is to get to a local minimum. As the “ball” rolls towards the minimum, 

it gains “momentum” and descends faster and faster. Once it overshoots, it will gain new 

momentum in the opposite direction, incentivizing it to reach the minimum faster).

There are more intricacies to the RBM, but in the end, you only need to know that 

RBMs can be used to create a probability distribution of the input data. We will use this 

property of RBMs to single out anomalies by checking the probability of that particular 

sample of occurring.