 |
|
|
|
|
|
|
|
|
|
|
|
|
 |
|
|||||||||||
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
|
|
 |
 |
|
||||||||||
 |
|
|||||||||||
 |
|
|||||||||||
 |
 |
|
||||||||||
| Seminar Schedule (Seminar Organizer: Prof. Ziv Bar-Joseph) | |
Machine Learning Lunchtime Chats
|
Fall 2001 Seminar
Date: December 3, 2001 Time: 12:30 PM - 2:00 PM Location: 3305 Newell-Simon Hall Speaker: Kevin Kelly Associate Professor, Philosophy, Carnegie Mellon University Title: Ockham's Razor Deduced from Error Minimization Abstract: Why should one choose the simplest theory if one doesn’t know in advance that reality is simple? In this talk, I define a structural concept of simplicity in terms of the branching structure of possible futures. Then I show that a taste for simplicity is necessary if inquiry is to efficiently converge to the right answer to a question, where efficiency is a matter of minimizing retractions or errors prior to converging to the right answer. The argument neither presupposes that the world is simple nor invokes prior probabilities biased toward simple worlds or hypotheses. I also show that in all efficiently solvable problems, a taste for simplicity is both necessary and sufficient for efficiency if efficiency is understood in terms of minimizing the commission of errors. If efficiency is understood in terms of minimizing retractions, a stronger condition is necessary and sufficient for efficiency: namely, that one never retract an answer unless a surprise or anomaly occurs. Finally, I show that a problem is efficiently solvable if it is solvable in the limit by a method that can also converge in the limit to “false presupposition” if the presupposition of the question is false. The results build upon a concept of transfinite retraction bounds introduced by computer scientists R. Freivalds and C. Smith.