Stability as performance metric for subjective pattern recognition - application of Electoral College in face recognition

For a class of pattern recognition problems, such as the face recognition problem, humans do not know the strategies that our brains employ in daily life and therefore there is no algorithm that can emulate our brain ability. Without understanding the psychological processes of brains, an objective of improving accuracy of such systems leads nowhere but to a trial-and-error process of different algorithms with different parameters. We argue that, at current stage, the objective of the research on such pattern recognition systems should be improving of the stableness/robustness rather than the accuracy. We further argue that stableness actually affects the accuracy. A stability model on voting systems has shown that Electoral College (aka regional vote scheme) is more stable than direct popular vote (aka national vote scheme), that the stability of Electoral College reaches the peak when the subdivided regions/windows are of middle size. By embedding regular face recognition algorithms into Electoral College framework, impressive performances have been achieved on several benchmark face sets.