Towards the Validation of Inductive Learning Systems

Abstract

The present paper deals with inductive inference of recursive functions, in general, and with the problem of validating inductive learning devices, in particular.

Thus, the paper aims at a contribution to the research and development area of intelligent systems validation. As those systems are typically interactive and, therefore, utilized in open loops of human-machine interactions, the problem of their validity is substantially complicated. A certain family of validation scenarios is adopted. Within this frame-work, we ask for the power and the limitations of these validation approaches. The expertise necessary and suficient to accomplish successful validation is of some particular interest. One of the key questions is for the comparison of domain expertise and validation expertise.

The area of inductive inference of recursive functions is taken as a case for complex interactive systems validation.

Computability theory is providing a rich source of theoretical concepts and results suitable for the focused investigations. Emphasis is put on explicating the importance of abstract computational complexity, limiting computability, and relativized computability. These concepts are exploited for characterizing the expertise necessary and suficient in the validation of inductive inference systems. Particular emphasis is put on relating validation expertise and domain expertise by means of relativized computability concepts. One of the key results on validation of inductive learning systems exhibits that validation expertise necessarily implies the expertise for solving the focused learning problems.