Engineerable AI Project
We envision and investigate “Engineerable AI” techniques to allow for tailing AI systems to meet strict requirements for dependable systems such as medical and automotive systems.
We focus on two core issues of machine learning techniques, especially deep learning, for AI systems: (1) dependency on enormous amount of data and (2) uncontrollably of the system behavior. Both of the issues are critical when we consider strict requirements for dependable systems such as medical and automotive systems. Regarding the first point, the functionality of AI systems, built from data, tends to be weak for specific minor cases. However, such minor cases are often significant to ensure safety or to complement human experts. Regarding the second point, fixing undesirable behavior or performance is difficult in a sense retraining based on data may change the whole behavior. This point leads to very high uncertainty and unpredictability in development, especially, continuous improvement, as retraining to fix something may break other part.
We have been investigating these issues by “Engineerable AI” techniques that allow for tailoring AI systems to meet strict requirements for dependable systems. On one hand, we work on AI construction techniques that allow for embedding human knowledge, thus avoiding relying only on data to discover significant properties of the target. On the other hand, we also work on AI debugging that allow to capture
We will demonstrate the “Engineerable AI” techniques with Proof-of-Concept studies on two types of quality-sensitive systems, medical systems and automotive systems.
eAI project is supported by the JST MIRAI funding scheme.