How does one design a mind? (In 4 billion years or less)

ghostslimΤεχνίτη Νοημοσύνη και Ρομποτική

23 Φεβ 2014 (πριν από 3 χρόνια και 5 μήνες)

87 εμφανίσεις

Human Research and Engineering Directorate

Troy Kelley

U.S. Army Research Laboratory

Human Research and Engineering Directorate

Aberdeen, MD

USA

How does one design a mind?

(In 4 billion years or less)

Human Research and Engineering Directorate

What is cognition?


Cognition is a collection of pre
-
programmed
algorithms developed during evolution


This is both high level


Language


Searching


And low level


Reflexes


Movement toward light

Human Research and Engineering Directorate

What is cognition? (con’t)


Cognition is also changes in neurological
connections based on experience


Learning at the low levels (reflexes)


And the high level as well (language)


If I know what cognition is, does that mean
I can recreate a cognitive system?

Human Research and Engineering Directorate


At the very least

a cognitive system needs:


Perceptual System


Visual


Auditory


Tactile


SICK, IR, LADAR


Memory System


LTM, STM, Working memory, visual spatial memory,
auditory memory (memory for each sensor?)


Hierarchical organization


Some kind of hierarchical organization process


Can’t really create a “black box”

Brain needs

Human Research and Engineering Directorate

Approaches to needs


Neurological Systems


Simulate every neuron


Symbolic Systems



Traditional AI systems


Complete sub
-
symbolic systems


Reactive architecture


Cognitive Architectures


ACT
-
R, Soar

Human Research and Engineering Directorate

Simulating Every Neuron

Source: Dr. Ray Kurzweil, Kurzweil Technologies

Approach


Neurological approach

Human Research and Engineering Directorate

How does Blue Gene, today’s most powerful
supercomputer, compare with the human brain?

*Data provided by Lawrence Livermore National Laboratory

Supercomputer and the Human Brain

Human brain is 100 times more powerful

Supercomputer


100,000 lbs


5,000 cubic ft


2,000,000 watts


100 trillion cycles

per second

Human Brain


4 lbs


0.06 cubic ft


?????


10 quadrillion cycles

per second

Human Research and Engineering Directorate

Approaches


Neurological Systems


Simulate every neuron?


How do we program all of those neurons?


Are they all basically the same or are they different?


We know from biological systems that different cells have
different functions even within the neurological system


So we can’t use one type of “perceptron” or neural network


Human Research and Engineering Directorate

Approaches


Neurological Systems


Simulate every neuron?


How do we program all of those neurons?


Are they all basically the same or are they different?


We know from biological systems that different cells have
different functions even within the neurological system


So we can’t use one type of “perceptron” or neural network


How do we determine the fitness of our cell clusters?



Human Research and Engineering Directorate

Charles Darwin Said….

“It is not the strongest of the species
that survives..…but rather the one
most responsive to change.”

Adaptation in Nature is essential!

Human Research and Engineering Directorate

Evolutionary approach


How to determine fitness?


Organisms

evolved
in conjunction

with
the
earth

evolving


Evolving a complex organism needs to be
done using a complex environment!

Human Research and Engineering Directorate

Approaches


Neurological Systems


Simulate every neuron?


How do we program all of those neurons?


Are they all basically the same or are they different?


We know from biological systems that different cells have
different functions even within the neurological system


So we can’t use one type of “perceptron” or neural network


How do we determine the fitness of our cell clusters?


Much of evolution has revolved around motor/sensor
optimization


is that the answer for robotics?



Human Research and Engineering Directorate

Sensor problem?


The creature with the best sensor wins?

Human Research and Engineering Directorate

Moth Sense and Control System


Biological sensors exhibit
unequaled sensitivity, specificity,
speed and refresh
-
rate



The chemical sensors of the moth can
detect a
single molecule

of the sex
pheromone of the female up to a mile
away

[Bazan lab, ICB, UCSB]


Signal amplification mediated by elements that fit together by
precise lock
-
and
-
key molecular recognition

Human Research and Engineering Directorate

Approaches


Neurological Systems


Simulate every neuron


Symbolic Systems


Traditional AI systems

Human Research and Engineering Directorate

AI Approach


Computationally intensive


Task specific


Not necessarily biologically based


Suffers from brittleness and lack of robust
behaviour in dynamic environments

Human Research and Engineering Directorate


"
In from three to eight years,

we'll have a
machine with the

general intelligence of an

average human being...


a machine that
will be able to

read Shakespeare [or]
grease

a car."



Marvin Minsky, Life magazine, 1970

AI answer

Approach


Traditional AI approach

Human Research and Engineering Directorate

Approaches


Neurological Systems


Simulate every neuron


Symbolic Systems


Traditional AI systems



Complete sub
-
symbolic systems


Reactive architecture

Human Research and Engineering Directorate

Reactive Architecture


Anti
-
symbolic


Tight pairing between sensing and
reaction


Current system for the military (4DRCS)


No representation of the environment

Human Research and Engineering Directorate

“Elephants Don’t Play Chess”


Rodney Brooks

Approach


Reactive Architecture

Humans do play chess, and perhaps we want to build robots that

can play chess

Human Research and Engineering Directorate

Approaches


Neurological Systems


Simulate every neuron


Symbolic Systems


Traditional AI systems


Complete sub
-
symbolic systems


Reactive architecture


Cognitive Architectures


ACT
-
R, Soar

Human Research and Engineering Directorate

Cognitive Architectures


Cognitive architectures have ignored the
“perceptual problem”


Cognitive architectures grew out of the
symbolic tradition of AI


Newell and Simon’s General Problem
Solver production system served as the
birth of AI as well as the birth of cognitive
architectures


Cognitive Architectures are complex

Human Research and Engineering Directorate

Complexity

A software mind should be at least as complex as an
operating system?



1993 Windows NT 3.1 6 million lines of code


1994 Windows NT 3.5 10 million lines of code


1996 Windows NT 4.0 16 million lines of code


2000 Windows 2000 29 million lines of code


2002 Windows XP 40 million lines of code



40 million lines of code and 9 years of development


Imagine this development cycle, except that, due to sensor error,
you never knew exactly where the user was clicking with the mouse,
or you never knew exactly what key was being selected on the
keyboard. How would this affect the development cycle?


Human Research and Engineering Directorate

Approaches


Neurological Systems


Simulate every neuron


Symbolic Systems


Traditional AI systems


Complete sub
-
symbolic systems


Reactive architecture


Cognitive Architectures


ACT
-
R, Soar


Hybrid approach


How do we merge a symbolic and sub
-
symbolic system?

Human Research and Engineering Directorate

Knowledge Architectures

Human Research and Engineering Directorate

Architectures for Modeling Cognition

X + Y = Z

Symbolic


Complex cognition

=

Serial in nature


Localized representation


Cognitive Architectures

Subsymbolic


Simple cognition

=

Parallel in nature


Distributed representation


Neural Networks

Human Research and Engineering Directorate

Intellectual continuum

within the human anatomy

Reflexes

The actions of reflexes are similar to a simple feed
-
forward

Neural Network

Frontal Lobes

The actions of the Frontal Lobes are similar to complex

Symbolic processing architectures

Kelley, T. D., (2003), “Symbolic and sub
-
symbolic representations in
computational models of human cognition: What can be learned from
biology?”
Theory and Psychology
, TAP 13(6), December.

Human Research and Engineering Directorate

Robotics Architectures


In

a

DARPA

report

(
2001
)

by

Singh

and

Thayer

of

the

CMU

Robotics

Institute

the

authors

concluded

that
:



“a

mixed

strategy

[hybrid]

provides

a

more

reasonable

method

for

robot

coordination

for

a

general

case

where

there

are

natural

constraints

during

operation

in

a

complex

environment
.



Human Research and Engineering Directorate

Stimuli

Subsymbolic processing

Production System

Goals

Camera inputs

Laser inputs

Sound inputs

Parallel processing

all of the inputs

simultaneously

Results go to memory

Production system

operates on memories

“Attention” is the highest

level goal

Semantic network

Human Research and Engineering Directorate

Sub
-
symbolic


How to develop pre
-
programmed
algorithms that look for one item?


Algorithms for corners, gaps, lines


Two programmers (graduate level)
working for one year


Still problems with these low level
algorithms


Human Research and Engineering Directorate


Human Research and Engineering Directorate

Conclusions


Complex behavior requires a complex
approach to cognition


Hybrid architectures offer one solution to a
complex problem


Combinations of symbolic and sub
-
symbolic architectures offer one approach