Conditional random fields

Discriminative models such as linear, logistic regression, multilayer perceptron, and support vector machines are described in part 3 – Gradient-based Learning. However, it would make sense to introduce a discriminative alternative to HMM in this chapter dedicated to sequential data models.

The conditional random field (CRF) is a discriminative machine learning algorithm introduced by John Lafferty, Andrew McCallum, and Fernando Pereira [7:9]. The algorithm was originally developed to assign labels to a set of observation sequences as found.

Let's consider a concrete example to understand the conditional relation between the observations and the label data.

Introduction to CRF

Let's consider the problem of detecting a fault during a soccer game using a combination of video and audio. The objective is to assist the referee and analyze the behavior of the players to determine whether an action on the field is dangerous (red card), inappropriate (yellow card), in doubt...