The Seminar on Dangerous Ideas

Gerald J. Sussman

PSBDARP or Learning Considered Dangerous

1 p.m. Wednesday, March 6, 2002

Abstract:

Very. Concrete:

I have been an observer and sometime participant in AI research since 1964. Over these many years I have seen several waves of work aimed at building systems that can be trained to perform novel tasks, not preprogrammed by their designers. Most attempts are based on abstraction of rules or on tuning of parameters. Neither approach has been particularly successful in approaching natural performance in the kind of tasks that separate humans from rodents.

I will address the reasons for the manifest failure of our attacks on the problem of constructing human-scale intelligence from the platform of learning. I believe that the failure is generic and predictable. I also believe that it is necessary that we repeatedly make these mistakes before we break through into a more fruitful path.

Learning Considered Harmful
or
PSBDARP

Problem
Solving
By
Debugging
Almost
Right
Plans

"call by purpose"
goal to compile a procedure
that answers to such a
goal -- using existing code.

Resting State
Smart people come in
"Universal" idea
Boom: desciples
Bust: idea is not so good
Deflation: cottage industry of "scientific" study
...and back to the resting state

Me interacts with World, sending Actuations and receiving Sensations. Inside Me is a Performer that does the interacting, and a Teacher that interacts with the performer. Most of what we do in machine learning is to tweak the messages the Teacher sends the Performer. We should take care that the language of these messages can (at least has the possibility to be able to) accomplish the desired change in the perfomer.