This is a live post from the World of Watson conference, hosted by IBM Watson.
[RF: This is live blog post. I strive accurately to report highlights of what I hear. Please forgive typos and anything I misunderstand. Comments in square brackets starting with RF are editorial comments and reflects solely my opinion. I occasionally insert these in live blog posts.] |
Live from IBM-hosted World of Watson, this is a live post covering.
Cognitive Computing in Perspective | Bruce Porter
Notes that we had AI surge in 1980s.Thousands of expert systems were built in the 1980s; they commonly performed as well as human experts. This was followed by winter of discontent in 1990s. Are we at risk for repeat. He thinks not.
To explain why, roll back the clock on AI. In 1955, John McCarthy defined the idea of building intelligent machines. Then, it was audacious, if not ridiculous. The first transitor based computers were built that year [RF: replacing vacuum tube ones].
Pioneers understood need for machine learning. First good example around 1960 was a machine that played checkers. Early AI researchers was psychologists at heart. They wanted to compare computing cognition to human cognition.
Problem with my account here and all such accounts, is that we miss the paths NOT taken. There was a competing strand: IA = Intelligence Amplification. Ross propoed in 1956 and Lichtlighter proposed a human-computer partnering in 1960. He anticipated many developments of today. But AI beat out IA.
The 1980s expert systems were idiot savants. They were great for an extremely narrow set of tasks but lacked the learning or flexibility to do more than extremely narrow tasks. The 80s expert systems bubble burst in early 90s. One reason is that it took much expert time to build. Another, is that they focused on the wrong domains: law and medicine. Of course, lawyers and doctors routinely did those tasks. The builders of the systems never asked them if they needed help. The problems solved for were the wrong ones.
AI languished in 1990s and had re-birth in 2000s. Probably a bit of a misperception and more of a shift from AI to IA. Now, focus is to aid humans, not replace.
Expert systems failed because it was too hard to tease out the information from the experts. Elucidating the rules by which experts reach conclusions is too hard. In contrast, machine learning (ML) does this automatically based on large volumes of examples. ML induces the rules from many systems.
A coming advance will be more dialog with system. Another will be discovery not just search. Says we could find cancer cures by learning from huge numbers of patterns – but augmented by human judgement.