Fisher’s Conditionality Principle in Statistical Pattern Recognition
Carey E. Priebe · The American Statistician · 2011
We present a simple, illustrative example of Fisher’s Conditionality Principle in statistical pattern recognition. We observe training data {(Xi, Yi, Zi)}i=1n with which to learn the discriminant boundary. At classification time, we observe the to-be-classified feature vector X with true-but-unobserved class label Y. We do not observe the Z associated with X, and the collection {Zi} is ancillary for the discriminant boundary. Nonetheless, {Zi} is essential for optimal classification.