Next: Extensions to the Basic Up: Basic Learning Rules Previous: Perceptron Contents


Naive Bayes

In the case of naive Bayes, a feature's weight in a target node is simply the logarithm of the fraction of positive examples (positive with respect to the target) in which the feature is also active (see, e.g., [Roth, 1999,Roth, 1998]). In addition, the relative weight of the target is used as ``prior'', and a fixed smoothing weight is added for evaluation of active features that were never observed in training4.7.

Notice that when using Perceptron and Winnow with the default training policy, all examples are presented to every target node, and an update may occur whether that target is active in the example or not. This is not the case for naive Bayes. Each target node takes into account only the examples labeled with it (i.e., the target's hypothesis is learned only from positive examples, and training is not mistake driven).

Naive Bayes does not apply a sigmoid function; its activation is equivalent to its sigmoid activation.



Footnotes

... training4.7
See the -b command line parameter for more information on smoothing in naive Bayes.


Next: Extensions to the Basic Up: Basic Learning Rules Previous: Perceptron Contents
Cognitive Computations 2004-08-20