Stochastic discrimination might appear to fall in the general category of pattern recognition methods based on the notion of combining classifiers. But there is a fundamental difference which epitomizes the essence of stochastic discrimination. Indeed, most combination methods deal with component classifiers which, even if they may be weak for the problem at hand, are still somewhat ``expert'' in that they will tend to agree with one another about the classification of at least some of the ``easy'' points in the feature space. However, in the case of stochastic discrimination, not only does this tend not to happen - it is essential to the success of the method that it be avoided to the greatest extent possible. We are not talking here about simply trying to maintain some degree of orthogonality among component classifiers. In the case of SD, all points of a given class in the feature space must be ``viewed equivalently'' by the full set of weak component classifiers. This notion, which in the theory of stochastic discrimination is known as ``uniformity,'' captures the essence of stochastic discrimination, making it, in effect, a method for combining ``classifiers'' which, modulo specific points in the feature space, have no commonality what-so-ever. In a sense, the weak component classifiers are not really classifiers at all - they are simply subsets of the feature space selected from a randomly generated stream of such subsets solely by virtue of their having an error rate for the problem at hand slightly better than the average for the stream.