A differential theory of learning for efficient statistical pattern recognition