Biometric Authentication: A Machine Learning Approach

3.5. Concluding Remarks

This chapter has detailed the algorithmic and convergence property of the EM algorithm. The standard EM has also been extended to a more general form called doubly-stochastic EM. A number of numerical examples were given to explain the algorithm's operation. The following summarizes the EM algorithm:

• EM offers an option of "soft" classification.

• EM offers a "soft pruning" mechanism. It is important because features with low probability should not be allowed to unduly influence the training of class parameters.

• EM naturally accommodates model-based clustering formulation.

• EM allows incorporation of prior information.

• EM training algorithm yields probabilistic parameters that are instrumental for media fusion. For linear-media ...