1/2 Criteria on AI Uncertainty Based on Shannon Entropy and Its Applications in Pattern Classification and Medical Imaging
Mingyong Zhou · Advances in transdisciplinary engineering · 2022
Classification problems are important in medical diagnosis. In this workshop, we present and summarize our recent insights into AI uncertainty through classification problems from both theoretical and empirical perspectives. First a concept and a theory are proposed aiming at zero-error AI system. We show in fact they can be derived from Shannon communication theory based on entropy concept. Classical Rademacher complexity and Shannon entropy is shown to be closely related by quantity by definitions. Based on this observation, we are able to derive a 1/2 criteria in terms of Shannon entropy that guarantee an AI zero-error accuracy in classifications problems. Last but not the least, we show both a relaxing condition and a stricter condition in real applications that can guarantee zero-error accuracy in AI classification problems. We provide examples to show in applications how to apply our derived 1/2 criteria into AI applications by “coding” methods.