Re:Sussman anomaly
by ryanianian on Monday January 15, 2018 @12:39PM
(#16151998)
> Machine learning is mostly about numerical optimization/exploration in high dimensions
That may be true right now, but eventually for ML to be "true AI" it will have to move to discrete optimization/exploration as opposed to (relatively) continuous problems like I think you're describing. This "anomaly" exists in the discrete space. The below conversation threads explore this better but with different words.
Or it will just get better and better at working with continuous approximations to discrete problems.
Re:Sussman anomaly
by ryanianian on Tuesday January 16, 2018 @10:54AM
(#16159052)
My intuition is that this is how the human brain works, but I think we'd be missing out on a big opportunity if we limit computers to this. A continuous solution to a discrete problem may hugely under-perform the truly optimized discrete solution.
A great deal of research was done (and is still being done) in discrete-reasoning AI; aside from a few successful applications (e.g., military logistics,) very little ever came of it, at least in the direction of "general intelligence."
Our intuitions about how our minds work are typically really bad, which makes it difficult to see what the needed techniques are to replicate them.
Machine learning is mostly about numerical optimization/exploration in high dimensions, so they can't really get stuck in the same way; your oscillating code could converge after a long time, but you probably need to solve the halting problem to know.