Lecture
Let be - the set of objects of the training sequence, that is, the belonging of each of them to one or another pattern is reliably known. Let also is the object closest to the recognizable . Recall that in this case the nearest neighbor rule for classification is that attributed to the class (image), which belongs . Naturally, this assignment is random. Probability that will be referred to there is a posteriori probability . If a very large, it is entirely possible to assume that located close enough to so close that . And this is nothing but a randomized decision rule: attributed to with probability . The Bayesian decision rule is based on the choice of the maximum a posteriori probability, that is, attributed to in case if
.
This shows that if close to unity, the nearest neighbor rule gives a solution, which in most cases coincides with Bayesian. Recall that these arguments have sufficient grounds only for very large (the volume of the training sample). Such conditions in practice are not common, but they allow one to understand the statistical meaning of the nearest-neighbor rule.
Comments
To leave a comment
Pattern recognition
Terms: Pattern recognition