Solve the AI Problem

boorishadamantAI and Robotics

Oct 29, 2013 (3 years and 11 months ago)

249 views

How We’re Going to

Solve the AI Problem

Pedro Domingos

Dept. Computer Science & Eng.

University of Washington




What is the AI Problem?

Build robots that do every job humans do,

as well as them or better.


(Preferably much better.)

Why Haven’t We Solved It Yet?


Because no one has really tried


Everyone works on subproblems


Because we don’t have the hardware


But this will soon change

A Phase Shift in AI Research


In 10 years (give or take), hardware will reach
the computational power of the human brain


Then we can really start trying


Progress is much faster when you work on
the actual problem


In the meantime: Lay the groundwork

Ways to Solve AI

Approach

Proponents

Examples

Mother of all KBs

Doug Lenat

Cyc

Web mining

Tom Mitchell

Oren Etzioni

WebKB

KnowItAll

Retrace evolution

Rod Brooks

Genghis, Cog

Robot baby

Paul Cohen

Robot Baby

One Algorithm

Geoff Hinton

Jeff Hawkins

Neural networks

Mother of All KBs


Hypothesis:

We don’t need no new discoveries;
just a lot of knowledge


Empirical test:

Miserable failure


It’ll take tens of thousands of rules …


No, hundreds of thousands …


No, wait, more like millions …


Deduction is not enough!


We need induction and uncertain reasoning


Cycorp now realizes this


… And at least they tried

Web Mining


Let’s read the Web instead of

manually inputting formal knowledge


Pros


There’s a lot of stuff in the Web


Language is great window into intelligence


Great application value in its own right


Cons


The Web sucks


Language is built on top of vision,

motor control, everyday life, etc.

Retracing Evolution


Human intelligence is too hard.

Build an insect first!


Well, that turns out to be easy, and doesn’t

bring us much closer to human intelligence


Brooks got tired after Genghis & pals,

and went straight to Cog (which did nothing useful)


Evolution is blindingly slow


Subsumption architecture still seems like

a good idea

Robot Baby


Build a robot and let it learn like a baby


Pros


Guaranteed to work! (Existence proof)


It solves the real problem


Cons


Is it overkill? (Intelligent
≠ Human)


Do we really have to wait 10 years for it to grow up?


Too much to try at once (start w. symbol grounding?)


And we don’t have the hardware …

One Algorithm


Hypothesis:

Neocortex is all one algorithm


Pretty good empirical support so far


It does everything: learning, reasoning, vision,

language, motor control, etc.


Shortest path to AI: Figure out what this algorithm is


Reverse engineer the brain? Not necessarily


Testbed: Digit recognition? No!


Algorithm has to work on many different problems
without change

How About This?


Build a
Robot Baby


Power it with
One Algorithm


Add stages one by one (
Subsumption
)


Feed it
Cyc


And then have it
Read the Web


It Takes a (Global) Village

[Richardson and Domingos, KCAP
-
2003]

Collective

Knowledge

Base

Users

Rules

Facts

Feedback

Queries

Answers

Outcomes

Inference

Learning

Contributors

What Can We Do Now?


Algorithms that work on any number of cores


Solve two problems simultaneously


Learning and reasoning


Vision and robotics


Language and common sense


Then solve three


Solve series of increasingly hard problems


Don’t get stuck in local optima


If you have 80/20 solution, move on to next harder problem


Stay off the bandwagons

Got the Hardware. Now What?


Divide and conquer doesn’t work for AI


Gluing pieces together doesn’t work

(engineering hits “complexity wall”)


We need the right
language


Mechanics:

Calculus


Electromagnetism:

Differential operators


Alternating current:

Complex numbers


Digital circuit design:

Boolean logic


AI:

Not there yet (but see Markov logic)

Three Simple Tests


You’re not solving the AI problem if …


Your system doesn’t work
online


Your system doesn’t
simultaneously

process
more than one type of information


Your system doesn’t process so much information

it needs a
focus of attention

mechanism


Consciousness
=
Lots of information

well
integrated

online

with a focus of attention

When Will We Solve AI?


Common view: Never


Kurzweil, Moravec: 25 years


Both wrong


Solving AI is a long
-
term project


How do we make sure we’re making progress?


How do we speed up progress?


How do we keep up motivation (and funding)?