Monday
May042020
Computers and Common Sense
Monday, May 4, 2020 at 2:31AM
This interesting article isn't about computer chess, but it is of possible relevance to it. Thoughts, especially from those of you who have some specialist knowledge about programming in general and computer chess programming in particular?
tagged computers
Reader Comments (4)
Fortunately, common sense isn't necessary to play chess well - it might even be a disadvantage. See: Bobby Fischer.
[DM: That's a different sort of common sense, or a different application - right though you may be. (Then again, trying to define a particular meaning of "common sense" would probably prove extraordinarily difficult.) One sort of chess example, which I've seen occur on multiple occasions, is a computer throwing away material - even massive material - to reach a tablebase ending.]
I actually work at the Allen institute for Brain Science. I'm a senior software engineer there and have done some forays into chess programming.
I don't want to make any claims about what the researcher being interviewed actually said but this is typical science journalism IMHO. Typical in that the author doesn't really understand what they are writing about very well And the researcher is really giving a bare bones explanation of things. Lots of misconceptions begin to sneak in. These errors are there because the journalist wants to explain the motivation for the later half of the article but the examples provided are just silly.
Honestly isn't very much interesting about It either not coming up with "fire" nor both coming up with fire and alternatives once you run it out of the only good answers. That tool was never built for common sense. IBM Watson was more designed in line with this It was quite good at little challenges like this a la it's Jeopardy appearance. When you use a hammer to bang in a screw you really shouldn't complain that the hammer is in a screwdriver. Especially when there are other tools like screwdrivers that can actually get the job done exactly the way you want.
There is a bit of a weak claim that neural networks are the solution to common sense but then of course is not true either. let's say that in your network really did mimic a neuron in the brain. That would be great but we couldn't actually create anything like a reasonable representation of a neuron population. It's just too big and too interconnected. But even if you could, common sense is not very likely to just spring forth because like all models they don't actually convey physical properties. Just like diagramming a black hole doesn't generate gravitational force, you really should not expect even a fully populated neural network model to generate a personality.
What's more is neural networks are still victims to their input data. The canonical issue that often is talked about is that you don't actually know why it makes the decisions it did (not exactly true but...). There is a popular tale about training and neural network to find tanks on the battlefield. It seemed to work but actually lost fidelity when the days got clearer. This is because it was actually training itself to, at least in part, look at clouds because that is a kind of data it was trained on.
Anyway the ability to create general purpose intelligence is very far away and these kind of fits and starts in the AI community usually end up being a discovery that leads to a very practical application followed by long periods of nothing until the next discovery. We are much better at creating specific purpose intelligence (also known as statistics ;) ) and finding ways to adapt own problems to look like problems that are special purpose intelligence machines could actually deal with. Finding little one off hiccups is almost never surprising or interesting.
Maybe when an Alpha Zero type engine like Leela or Fat Fritz has played a billion or even a trillion games and can use that knowledge will computers show some sign of common sense. When a 1800+ player can see that a position is a dead draw but not a 3500+ chess engine, we know chess engines cannot think. Their skills still rely on the programming that teaches them how to compute numbers.
Ross's comment was very thought provoking. It got me thinking that human brains also do not always process information in intelligent and productive ways. So when the Brain Scientists are trying to simulate a human brain , just what does that really mean. Just for an example Robert Fischer's brain produced some the finest chess ever by a human, but his brain also lead him to believe some of the most disturbing ideas about human behavior. I don't mean to pick on Fischer, he is just an, chess player, example of the dichotomy often found in genius. Fischer seemed to lack common sense in his social views but showed a near perfect logic in his chess play.
Maybe when we evaluate machines that "think," we need to have an understanding of why great intelligence and poor judgement are part of the same intellect.