人工智能簡介

ANINTRODUCTION

TOARTIFICIAL

INTELLIGENCECOMPILED

BY

HOWIE

BAUM12Artificial

intelligence

(AI),

sometimes

called

machine

intelligence,is

intelligence

demonstrated

by

machines,

in

contrast

to

the

naturalintelligence

displayedby

humans

and

other

animals,

such

as

"learning"and

"problem

solving.

.In

computerscienceAI

research

is

defined

as

the

study

of"intelligent

agents":any

device

that

perceives

its

environment

andtakes

actions

that

maximize

its

chance

of

successfully

achievingitsgoals.HOW

ARE

HUMANS

INTELLIGENT

?LearningReasoningProblem

Solving

and

CreativitySocial

BehaviorExperiencing

our

Environment

with

our

senses:HearingSightTouchTasteSmelling345Ways

that

People

Think

and

LearnAbout

ThingsIf

you

have

a

problem,

think

of

a

past

situationwhere

you

solved

a

similar

problem.If

you

take

an

action,

anticipate

what

might

happennext.If

you

fail

at

something,

imagine

how

you

mighthave

done

thingsdifferently.If

you

observe

an

event,

try

to

infer

what

prior

eventmight

have

caused

it.If

you

see

an

object,

wonder

if

anyone

owns

it.If

someone

does

something,

ask

yourself

what

theperson's

purpose

was

in

doingthat.This

is

whatHumans

dobestCan

youlistthe

itemsinthis

picture

?Acomputermight

havetroubleidentifying

thecatthere.Can

you

count

thedistribution

ofletters

in

a

book?Add

a

thousand4-digit

numbers?Match

fingerprints?Search

a

list

of

amillion

values

for

duplicates6?For

example,

writing

a

program

to

pick

out

objects

in

a

picture:This

is

whatComputers

do

bestArtificial

intelligence(AI)

-

The

study

of

computer

systemsthatattempt

to

model

and

apply

the

intelligence

of

the

human

mind.When

we

compare

Humans

to

Machines,

it

is

important

to

note

that

aMachine

can

be

a

car,

a

Smart

Phone,

a

Digital

Television,

etc.789The

illustration

below

illustrates

a

typical

information

flow

between

the"human"

and

"machine"

components

of

a

system. For

a

properly

designedsystem,

its

important

to

know

the

capabilities

and

flexibilities

of

both./webtraining/HFModel/HFInterModel/overview.htm10KEY

RESEARCH

AREAS

IN

AIProblem

solving,

planning,

and

search

---

generic

problem

solvingarchitecture

based

on

ideas

from

cognitive

science

(game

playing,robotics).Knowledge

Representation

–

to

store

and

manipulate

information(logical

and

probabilistic

representations)Automated

reasoning

/

Inference

–

to

use

the

stored

information

toanswer

questions

and

draw

new

conclusionsMachine Learning

–

intelligence

from

data;

to

adapt

to

newcircumstances

and

to

detect

and

extrapolate

patternsNatural

Language

Processing

–

to

communicate

with

the

machineComputer

Vision

---

processing

visual

informationRobotics

---

Autonomy,

manipulation,

full

integration

of

AIcapabilities1112From

SIRI

and

Alexa,

to

self-driving

cars,

artificialintelligence

(AI)

is

progressing

rapidly.While

science

fiction

often

portrays

AI

as

robots

with

human-likecharacteristics,

AI

can

encompass

anything

from

Google’s

searchalgorithms,

to

IBM’s

Watson,

to

autonomous

weapons.Artificial

intelligence

today

is

properly

known

as

narrow

AI

(or

weak

AI),

in

that

it

is

designed

to

perform

a

narrow

task

such

as

only

facial

recognition,

or

only

internetsearches,

or

only

driving

a

car).However,

the

long-term

goal

of

many

researchers

is

tocreate

general

AI

(AGI

or

strong

AI).While

narrow

AI

may

outperform

humans

at

whatever

itsspecific

task

is,

like

playing

chess

or

solving

equations,

AGIwould

outperform

humans

at

nearly

every

thinking

task.1314The

potential

benefits

from

self-learning

computer

chips

arelimitless

as

these

types

of

devices

can

learn

to

perform

the

mostcomplex

thinking

tasks,

such

as

interpreting

critical

cardiacrhythms,

detecting

anomalies

to

prevent

cyber-hacking

andcomposing

music.This

is

a

new

one

made

by

the

Intel

company

and

many

othercompanies

are

making

special

AI

chips

too.AUTOMATONS

–

ARE

THESE

DEVICESINTELLIGENT

?/watch?v=C7oSFNKIlaM

(2.22

min)15Artificial

Intelligence

(AI)

has

entered

our

daily

lives

like

never

beforeand

we

are

yet

to

unravel

the

many

other

ways

in

which

it

could

flourish.All

of

the

tech

giants

such

as

Microsoft,

Uber,

Google,

Facebook,

Apple,Amazon,

Oracle,

Intel,

IBM

or

Twitter

are

competing

in

the

race

to

leadthe

market

and

acquire

the

most

innovative

and

promisingAIbusinesses.16171819/watch?v=GoXp1leA5QcGoogle

announced

their

Duplex

system,

a

new

technology

forconducting

natural

conversations

to

carry

out

“real

world”

tasks

overthe

phone.The

technology

is

directed

towards

completing

specific

tasks,

suchas

scheduling

certain

types

of

appointments.For

such

tasks,

the

system

makes

the

conversational

experience

asnatural

as

possible,

allowing

people

to

speak

normally,

like

theywould

to

another

person,

without

having

to

adapt

to

a

machine.2021/watch?v=gsUV0mGEGaY22232425The

answer

is

all

of

the

above.Each

of

these

highly

realistic

images

were

created

by

generativeadversarial

networks,

or

GANs.GAN,

a

concept

introduced

by

Google

researcher

Ian

Goodfellow

in2014,

taps

into

the

idea

of

“AI

versus

AI.”There

are

two

neural

networks:

the

generator,

which

comes

up

with

a

fake

image

(say

a

dog

for

instance),

and

a

discriminator,which

compares

the

result

to

real-world

images

and

gives

feedbackto

the

generator

on

how

close

it

is

to

replicating

a

realistic

image.262728Turing

testA

test

to

determine

whether

a

computer

has

achieved

intelligenceAlan

TuringAn

English

mathematician

who

wrote

a

landmark

paper

in

1950

thatasked

the

question:

Can

machines

think?He

proposed

a

test

to

answer

the

question

"How

will

we

know

whenwe

have

succeeded?“He

said

that

a

machine

passes

the

test

when

it

successfully

generatesresponses

appropriate

enough

to

convince

the

evaluator

that

it

ishuman.29The

Turing

TestIn

the

Turing

test,

the

interrogator

must

determine

whichrespondent

is

the

computer

and

which

is

the

human.3031THE

LOEBNER

PRIZE

FOR

COMPLETING

THE

TURING

TESTThe

Loebner

Prize

is

an

annual

competition

in

artificialintelligence

that

awards

prizes

to

the

computerprograms

considered

by

the

judges

to

be

the

most

human-like,

usingthe

TuringTest

computer

and

person

arrangement.The

contest

was

launched

in

1990

by

Hugh

Loebner

and

thereare

bronze,

silver,

and

gold

coin

prizes,

plus

money.So

far,

there

have

only

been

winners

of

the

bronze

medal

and

a$4,000

award.32Silver

–aone-time-only

prize

plus

$25,000

offered

for

the

firstprogramthat

judges

cannot

distinguishfrom

a

real

human.Gold

plus$100,000

for

the

firstprogram

that

judges

cannot

distinguish

froma

real

human

in

a

Turing

test

that

includes

deciphering

and

understandingtext,

visual,

and

auditory

input.Once

this

is

achieved,

the

annual

competition

will

end..KNOWLEDGE

REPRESENTATIONWe

need

to

create

a

logical

view

of

the

data,

based

on

how

we

wantto

process

itNatural

language

is

very

descriptive,

but

does

not

lend

itself

toefficient

processing.What

are

the

different

ways

that

we

can

represent

knowledgeso

itcan

be

reviewed

by

an

Artificial

Intelligence

computerprogram

?Expert

LearningSystemsSemantic

Networks

-

A

knowledge

representation

technique

thatfocuses

on

the

relationships

and

word

descriptions

of

objects. A

graphis

used

to

represent

a

semantic

network

or

netDecision

or

Search

treeNeural

networks

–

creating

a

computer

version

of

the

neurons

of

thebr3a3inandhowthey

work12-341) Expert

Learning

SystemsExpert

Learning

Systems

were

commercially

the

first

and

mostsuccessful

domain

in

Artificial

Intelligence.Somewhat

out

of

favor

todayThese

programs

mimic

the

experts

in

whatever

field

is

beingstudied.Auto

mechanicCardiologistOrganic

compoundsMineral

prospectingInfectious

diseasesDiagnostic

internal

medicinecomputer

configurationEngineering

structural

analysisAudiologistTelephone

networkingDelivery

routingProfessional

auditorManufacturingPulmonary

functionWeather

forecastingBattlefield

tacticianSpace-station

life

supportCivillawRule-based

or

Expert

systems

-

Knowledgebasesconsisting

of

hundreds

or

thousands

of

rules

of

theform:IF

(condition)

THEN

(action).Use

rules

to

store

knowledge

(“rule-based”).The

rules

are

usually

gathered

from

experts

in

the

field

beingrepresented

(“expert

system”).Most

widely

used

knowledge

model

in

the

commercial

world.IF

(it

is

raining

AND

you

must

go

outside)THEN

(put

on

your

raincoat)Rules

can

fire

off

a

chain

of

other

rulesIF

(raincoat

is

on)THEN

(you

will

not

get

wet)Expert

Systems36GardenerExpert

System

ExampleExpertSystemsNamed

abbreviations

that

represent

conclusions:NONE—apply

no

treatment

at

this

timeTURF—apply

a

turf-building

treatmentWEED—apply

a

weed-killing

treatmentBUG—apply

a

bug-killing

treatmentFEED—applya

basic

fertilizer

treatmentWEED

&

FEED—apply

a

weed-killing

and

fertilizer

combinationtreatment3738ExpertSystemsVariables

that

are

needed

to

represent

thestate

ofthelawnBARE—the

lawn

has

large,

bare

areasSPARSE—the

lawn

is

generally

thinWEEDS—the

lawn

contains

many

weedsBUGS—the

lawn

shows

evidence

of

bugs39Expert

SystemsData

that

is

available:LAST—the

date

ofthelast

lawn

treatmentCURRENT—current

dateSEASON—the

current

seasonNow

we

can

formulate

some

rules

forourgardening

expert

systemRules

take

the

form

ofif-then

statements40ExpertSystemsSomerulesif

(THE

CURRENT

DAY

–

LAST

DAY

IS

LESS

THAN

30)

thenNONEif

(SEASON

=

winter)

then

not

BUGSif

(BARE)

then

TURFif

(SPARSE

and

not

WEEDS)

then

FEEDif

(BUGS

and

not

SPARSE)

then

BUGif

(WEEDS

and

not

SPARSE)

then

WEEDif

(WEEDS

and

SPARSE)

then

WEED

&

FEED41Expert

SystemsAn

execution

of

our

inference

engineSystem:User:Does

the

lawn

have

large,

bare

areas?NoSystem:User:Does

the

lawn

show

evidence

of

bugs?NoSystem:User:Is

the

lawngenerally

thin?YesSystem:User:Does

the

lawncontain

significant

weeds?YesSystem:Youshouldapply

a

weed-killing

and

fertilizercombination

treatment.422)

Semantic

(word

description)

NetworksSemantic

networkA

knowledge

representation

technique

that

focuses

on

therelationships

between

objectsA

directed

graph

or

word

chart

is

used

to

represent

a

semanticnetwork

or

net3)

Search

TreesAI

often

revolves

around

the

use

of

algorithms.An

algorithm

is

a

set

of

instructions

that

a

mechanical

computercanexecute.A

complex

algorithm

is

often

built

on

top

of

another,

simpler,

oneand

a

common

way

tovisualize

itis

with

a

treedesign.43A

simple

example

of

an

algorithmisthe

following

recommendations

foroptimal

play

at

tic-tac-toe:If

someone

has

a

"threat"(that

is,two

in

a

row),

take

the

remainingsquare.

Otherwise,Ifa

move

"forks"

to

create

twothreats

at

once,

play

that

move.Otherwise,Take

the

center

square

if

it

isfree.Otherwise,If

your

opponent

has

played

in

acorner,

take

the

opposite

corner.Otherwise,Take

an

empty

corner

if

one

exists.Otherwise,Take

any

empty

square.4445An

example

is

a

Search

tree

for

playing

the

game

Tic-Tac-Toe,

as

shownbelow.This

image

depictsmany

of

the

possiblepaths

that

the

game

can

takefromthe

having

the

first

2

rows

filled,asshown:THEHUMANBRAINANDNEURONSIN

ITA

REVIEW

BEFORE

THE

DISCUSSION

ABOUT4)NEURALNETS46THE

BRAIN

IS

DIVIDED

INTO

4LOBES

AND

THE

CEREBELLUMWHICH

IS

LOCATED

AT

THEBOTTOM,

BACK

AREA47AI

technology

called

machine

learning

today,

is

great

at

helping

fortaking

good

photos,

translating

languages,

recognizing

your

friendson

Facebook,

delivering

search

results,

screening

out

spam

and

manyotherchores.It

usually

uses

an

approach

calledneural

networks

that

workssomething

like

a

human

brain,

not

a

sequence

of

IF

THIS,

THEN

stepsas

in

traditional

computing.48TYPES

AND

FUNCTION

OF

NEURONSNeurons

are

essential

for

every

action

that

our

body

and

brain

carry

out.It

is

the

complexity

of

neuronal

networks

that

gives

us

our

personalities

andourconsciousness.They

make

up

around

10

percent

of

the

brain;

the

rest

consists

of

glial

cellsand

other

cells

that

support

and

nourish

the

neurons.49Thereare

around

86

billion

neurons

in

the

brain. To

reach

this

hugetarget,

a

developing

fetus

must

createaround

250,000

neuronsperminute

!Each

neuron

is

connected

to

at

least

10,000

others

–

giving

well

over1,000

trillion

connections

(1

quadrillionconnections).They

all

connect

at

a

junction

called

a

synapse,

which

can

be

electrical

or

ahigherpercentage

of

them

are

chemical.50Incoming

signals

to

the

neuron

can

be

either

excitatory

–

which

means

theytend

to

make

the

neuron

fire

(generate

an

electrical

impulse)

–

or

inhibitory

–which

means

that

they

tend

to

keep

the

neuron

from

firing.A

single

neuron

may

have

more

than

one

set

of

dendrites,

and

may

receive

manythousands

of

input

signals.Whether

or

not

a

neuron

is

excited

into

firing

an

impulse

depends

on

the

sum

ofall

of

the

excitatory

and

inhibitory

signals

it

receives.If

the

neuron

does

end

up

firing,

the

nerve

impulse

is

conducted

down

the

axon.51https://www.youtu

/watch?v=m

ItV4rC57kM5&2

t=10sHow

synapses

work

-

Neurons

are

connectedtoeach

other

at

alocationcalled

a

Synapse,

so

that

they

can

communicate

messagesAmazingly,

where

each

cell

connects

with

theother

one,

NONE

ofthesecells

ever

touch

each

other

!!The

signal

thatis

carried

from

thefirst

nervefiberto

the

next

oneistransmitted

by

an

electrical

signal

or

a

chemical

one,

up

to

a

speedof

268

miles

perhour!There

is

new

evidence

that

both

types

closely

interact

with

each

other

andthat

the

transmission

of

a

nerve

signal

is

both

chemical

and

electrical,which

is

actually

required

for

normal

brain

development

and

function.53If

you

don’t

use

a

foreign

language

you

learned

years

ago

ormathematics,

the

neurons

used

for

those

things

will

move

thesynapses

away

from

each

other

so

they

can

do

other

things

thatyou

are

learning

to

do. This

is

called

Synaptic

Pruning.544)

Artificial

Neural

Network

(ANN)A

computer

representation

of

knowledge

that

attempts

tomimic

the

neural

networks

of

thehuman

brainYes,

but

what

is

a

human

neural

network?Neural

networks,

or

neural

nets,

were

inspired

bythearchitecture

of

neurons

in

thehumanbrain.A

simple

"neuron"

N

accepts

input

from

multiple

otherneurons,

each

of

which,

when

activated

(or

"fired"),

cast

aweighted

"vote"

for

oragainst

whether

neuron

N

should

itselfactivate.55An

ANN

is

based

on

a

collectionof

connected

units

or

nodes

called

artificial

neurons,

which

loosely

model

the

neurons

in

a

biological

brain.Each

connection,

like

the

synapses

in

a

biological

brain,

can

transmit

a

signal

from

one

artificial

neuron

to

another.

An

artificial

neuron

that

receivesa

signal

can

process

it

and

then

signal

additional

artificialneurons

connected

toit.12-56ARTIFICIAL

NEURAL

NETWORKArtificial

neurons:

Commonly

called

processing

elements,are

modeled

after

real

neurons

of

humans

and

other

animals.Has

many

inputs

and

one

output.The

inputs

are

signals

that

are

strengthened

or

weakened(weighted).If

the

sum

of

all

the

signals

is

strong

enough,

the

neuronwill

put

out

a

signal

to

the

next

neuron

output

of

a

1.OutputArtificialNeuronInputs57Artificial

Neural

NetworksTrainingThe

process

of

adjusting

the

weights

and

threshold

values

in

aneural

netHow

does

this

all

work?Train

a

neural

net

to

recognize

An

eagle

in

a

picture.Given

one

output

value

per

pixel,

train

network

to

produce

anoutput

value

of

1

for

every

pixel

that

contributes

to

the

eagle

and0for

every

one

that

doesn’t.58DeepMind

is

a

subsidiary

of

Google

that

focuses

on

thedevelopment

of

artificial

intelligence

and

deep

reinforcementmachine

learning.The

deep

reinforcement

learning

of

its

AI

algorithms

has

beenused

in

both

research

and

applied

contextsDeepMind

is

built

around

the

framework

of

neural

networks

anduses

a

method

called

deep-reinforced-learning.This

means

that

the

A.I

can

learn

from

it's

experiences

andbecome

more

efficient

at

whatever

it

does.The

A.I

is

general-purpose

meaning

that

it's

NOT

pre-programmedfor

a

specific

task

from

the

go./watch?v=gn4nRCC9TwQ59AgentsAn

agent

is

anything

that

can

be

viewed

as

a

device

thatcan

perceive

its

environment

through

sensors

and

act

uponthat

environment

through

actuators.Human

agent:

eyes,

ears,

and

other

organs

for

sensors;

hands,legs,

mouth,

and

other

body

parts

for

actuatorsRobotic

agent:

cameras

and

infrared

range

finders

for

sensorsVarious

motors

for

actuatorsRational

Agent:For

each

possible

sequence,

a

rational

agent

should

select

anaction

that

is

expected

to

maximize

its

performance

measure,given

the

evidence

provided

by

the

perception

sequence

andwhatever

built-in

knowledge

the

agent

has.Why

“meaning”is

the

central

concept

of

AIFor

an

agent

to

be

“intelligent”,

it

must

be

able

to

understand

themeaning

of

information.Information

is

acquired

/

delivered

/

conveyed

in

messages

whichare

phrased in

a

selected

representation

language.There

are

two

sides

in

information

exchange:

the

source

(text,image,

person,

program,

etc.)

and

the

receiver

(person

or

an

AIagent).

They

must

speak

the

same

“l(fā)anguage”

for

the

informationto

be

exchanged

in

a

meaningful

way.The

receiver

must

have

the

ability

to

interpret

the

informationcorrectly

according

to

the

intended

by

the

source

meaning

or

semantics

ofit.MEANING

=

SEMANTICS606162Machine

LearningThe

phrase

‘machine

learning’

dates

back

to

the

middle

of

the

lastcentury

where

Arthur

Samuel

in

1959

defined

machine

learning

as“the

ability

to

learn

without

being

explicitly

programmed.”Machine

learning

is

a

type

of

AI

that

helps

a

computer’s

ability

tolearn

and

essentially

teach

i