4 ms·
What Is Explainable AI and What Is It Used For?
- nonrandomstring 5y agoWe had this in the 80s and called it "introspection". The problem with using the bogus term "AI" (other than that it isn't intelligent) is we further conflate wildly disparate technologies capable of solving problems based on prior input. Procedural systems I worked with were called "Expert systems" and had amazing "blackboard walking" capabilities, much like a stack trace. "sel X because X is member of Y (p=0.784) -> X fails crit Z -> search-new X -> fail: backtrack select Y" etc... You could learn a lot about the problem space and discover new facts and relationships by talking to the system about what it is doing and why. That's actually huge fun! The NIST report in TFA looks fanciful and some of the problems listed are hard, even in procedural systems that have explicit state and logic. For example, a system "knowing its limits". This seems in the order of Russell and Whitehead, as it isn't even clear that humans understand how to "know our limits". Indeed "AI" may be greatest symptom of that. What I see instead is "Assurance", and to me, assurance is generally bulllshit. Introspection of neural systems is going to be fuzzy and require further "AI" systems to interpret and humans to sugar coat. We will still need expert witnesses to toil though the output in courts. However, the most sinister line in TFA was this: "to help systems gain trust and acceptance in society" Now we wish to build systems with the a-priori aim of insinuating their trust and figuring out how to make themselves "acceptable"? We are literally moving toward programmatic psychopathy. Ironic. When it comes to "knowing limits" how is it we don't see that there are some combinations of computing and human affairs that should not be attampted.