每位技術(shù)專家都應(yīng)了解的人工智能和深度學(xué)習(xí)要點(diǎn)

What

Every

TechnologistShould

Know

About

AI

andDeepLearningAlex

McDonaldStandards

&

Industry

Associations,

NetApp

Inc.The

information

is

intended

to

outline

our

general

product

direction.

It

isintendedfor

information

purposesonly,and

may

not

be

incorporated

into

any

contract.

It

isnot

a

commitment

to

deliver

any

material,

code,

or

functionality,

and

should

notbe

relied

upon

in

making

purchasing

decisions.

NetApp

makes

no

warranties,expressed

or

implied,

on

future

functionality

and

timeline.

The

development,release,

and

timing

of

any

features

or

functionality

described

for

NetApp’sproducts

remains

at

the

sole

discretion

of

NetApp.

NetApp's

strategy

andpossible

future

developments,

products

and

or

platforms

directions

andfunctionality

are

all

subject

to

change

without

notice.

NetApp

has

no

obligation

topursue

any

course

of

business

outlined

in

this

document

or

anyrelatedpresentation,

or

to

develop

or

release

any

functionality

mentioned

therein.2Why

is

AI

&

Deep

Learning

Important?AI

and

Deep

Learning

is

disrupting

everyindustryFor

decades,

AIwas

allabout

improvingalgorithmsNow

thefocus

ison

putting

AI

topracticaluseCritical

to

leverage

well-engineeredsystemsThis

talkwillTake

you

on

a

broad

coherent

tour

of

Deep

Learning

systemsHelp

you

appreciate

the

rolewell-engineered

systems

play

in

AIdisruptionTake

you

a

step

closer

to

being

a

unicornSystems

+

AI

-

something

highly

desirable,

difficult

to

obtainNetApp

INSIGHT

?2019

NetApp,

Inc.

All

rightsreserved.

NetApp

Confidential–

Limited

Use

Only3AgendaAI

PrimerAI

Stacks

OverviewDeep

Learning

ProcessTrainingInferenceDeep

Learning

SystemsHardwareSoftwareDatasets

and

DataflowFutureof

Systems4BackgroundAI

or

ML

orDLAI

–

program

that

imitates

human

intelligenceML

–

program

that

learns

with

experience

(i.e.,

data)DL

–

ML

using

>1

hidden

layers

of

neural

networkNetApp

INSIGHT

?2019

NetApp,

Inc.

All

rightsreserved.

NetApp

Confidential–

Limited

Use

Only5Deep

Learning

101Basic

concepts

and

terminologyNeuron:

computational

unitDL

Model==type&

structureMore

layers

=>

better

capture

the

features

indataset,

better

performance

at

task

(normally)Parameters/WeightsNeuronParameter/WeightsLayer6Deep

Learning

101Basic

concepts

and

terminologyNeuron:

computational

unitDL

Model==type&

structureMore

layers

=>

better

capture

the

features

indataset,

better

performance

at

task

(normally)Parameters/WeightsTraining:

build

a

model

from

datasetEpoch:

a

pass

over

entire

datasetBatch:

a

chunk

of

dataPreprocessing/preparation:

ready

data

to

trainBackpropagationRepeatForward

propagation“mountain”data7Deep

Learning

101Basic

concepts

and

terminologyNeuron:

computational

unitDL

Model==type&

structureMore

layers

=>

better

capture

the

features

indataset,

better

performance

at

task

(normally)Parameters/WeightsTraining:

build

a

model

from

datasetEpoch:

a

pass

over

entire

datasetBatch

size:

a

chunk

of

dataPreprocessing/preparation:

ready

data

to

trainInference:

usinga

trainedmodelForward

propagation“mountain”New

data8Deep

Learning

101Basic

concepts

and

terminologyForward

propagationState-of-the-Art

DL

is

large

scale100s

of

layers

Millions

ofparameters100s

of

GBs

to

TBs

of

dataHours/days

to

train“mountain”data9AI

StackOverview10AI

StackLayersAIPaaS,

End-to-end

solutionsAIStackLayersGPUs,

TPUs,

FPGAsOptimized

hardware

to

provide

tremendousspeed-up

for

training,

sometimesinferenceMore

easily

available

on

cloud

for

rentModern

Compute11AIStackLayersTensorflow,

PyTorch,

Caffe2,

MxNet,

CNTK,Keras,

GluonLibrary

that

implements

algorithms,

providesexecution

engine

andprogramming

APIsUsed

to

train

and

build

sophisticated

models,and

to

do

predictions

based

on

the

trainedmodel

for

new

dataSoftwareModern

Compute12AIStackLayersLaptop,

Cloud

compute

instances,

H2O

Deep

Water,

Spark

DL

pipelinesHardware

accelerated

platforms,

supportingcommon

software

frameworks,

to

run

thetraining

and/or

inference

of

deep

neuralnetworksTypically

optimized

for

a

preferred

softwareframeworkCan

be

hostedon-prem

or

cloudAlso

offered

as

fully-managed

service

(PaaS)by

cloud

vendors

like

Amazon

SageMaker,Google

Cloud

ML,

Azure

MLPlatformSoftwareModern

Compute13AIStackLayersAmazon

Rekognition,

Lex

&

Polly;

GoogleCloud

API;

Microsoft

Cognitive

Services;Allows

query

based

service

access

togeneralizable

state

of

art

AI

models

forcommontasksEx:

send

an

image

and

get

object

tags

as

result,

send

mp3and

get

converted

text

as

result

and

so

onNo

dataset,

no

training

of

model

required

by

userPer-call

cost

modelIntegrated

with

cloud

storage

and/or

bundled

intoend-to-end

solutions

and

AI

consultancyofferingslike

IBMServices

Watson

AI,

ML

&

Cognitive

consulting,

Amazon's

ML

Solutions

Lab,Google's

Advanced

Solutions

LabAPI-based

servicePlatformSoftwareModern

Compute14Deep

LearningProcessTrainingInferenceDL

Process

and

Data

LifecycleDL

lifecycle

is

very

unlike

traditional

systems

software

developmentGather

DataData

Analysis,Transformation,ValidationModel

TrainingModelEvaluation,Validation,

TuningModelServing,MonitoringGathering

and

curating

quality

datasets

andmaking

them

accessible

across

orgDiverse

tools

and

flexible

infrastructure

neededEvaluation

criteria

is

criticalbut

hardComparing

algorithms

is

not

straightforwardTracking

artifacts

like

dataset

transformations,tuning

history,

model

versions,

validation

resultsmore

important

thancodeDebugging,

interpretability

and

fairness

islimitedTension/friction:

Data

security

and

privacy;

ITNetApp

INSIGHT

?2019

NetApp,

Inc.

All

rightsreserved.

NetApp

Confidential–

Limited

Use

Only16Deep

Learning

TrainingTraining:

build

a

model

from

a

datasetIs

memory

and

computeboundBig

datasets,

complex

math

operationsIs

highly

parallelized/distributed

–

acrosscores,across

machinesPartition

data,

or

model,

or

bothScale

Up

before

Scale

OutCommunication

to

computation

ratioSpeed

vs

Accuracy

tradeoffFederated

learningLeans

on

enhancements

to

data

qualityAugmentation,

randomnessTransformationsEfficiently

fit

in

memoryForward

propagationBackpropagationRepeat“mountain”17Deep

Learning

TrainingTraining:

build

a

model

from

a

datasetSupervision

-

rely

on

labeled

dataTransfer

learning:

a

pre-trained

model,

train

few

layersLearning

labelInvolves

a

lot

of

hyperparameter

tuningExample:

#layers,

#neurons,

batch

size

…Multi-model

training

on

same

data

setTrial

and

error

search

-

easier

to

automateRise

of

AutoMLLearn

how

to

model

given

data

–

nomodeling/tuningexpertise

requiredExample:

AmoebaNet

beat

ResNet

ImageClassificationForward

propagationBackpropagationRepeat“mountain”18Deep

Learning

InferenceInference*:

use

a

trained

model

on

new

dataIs

computationally

simplersingle

forward

passTypically

a

containerized

RPC/Web

serverwith

pre-installed

DL

software

+

NNmodelMultiple

inputs

are

batched

for

better

throughputBut

much

smaller

than

training

batchLowlatency*

aka

Model

Serving,

Deployment,

PredictionForward

propagation“mountain”19Deep

Learning

InferenceInference*:

use

a

trained

model

on

new

dataDL

models

can

be

hugemay

need

hardware

accelerationOn-device/Edge

inference

is

gaining

tractionReason:

latency

&

privacyspecial

modeloptimizations

–pruninghardware

on-devicePortability

and

interoperability

of

model

isimportantTrain

any

way,

deploy

anywhereExample:

ONNX

is

a

step

towardsstandardizing*

aka

Model

Serving,

Deployment,

PredictionForward

propagation“mountain”20Deep

LearningSystemsHardware

AccelerationSoftware

FrameworksCPUs

are

still

used

for

ML

trainingCPUs

are

common

forinferenceincluding

certain

DL

inferenceStruggle

to

handle

DL

trainingData

preprocessing

are

suited

for

CPUsHybrid

hardware

of

CPUs

with

otheraccelerators

is

commonRole

of

CPU

in

AIAcceleratorslike

GPU22GPU/Image

source:

https://.

Courtesy

Daniel

Whitenack.De

facto

hardware

for

AI

trainingAlso

for

large

scaleinferenceGPU

vs

CPU

:

many

more

cores,

parallelizationModern

GPU

architectures

used

for

AIHigh

speed

interconnect

between

CPU/GPUs(NVLink)Bypass

CPU

for

communication

(GPUDirect)Efficientmessage

passing

(Collective-All-Reduce)Available

in

cloud

(EC2

P*

instances)

and

on-premise

(DGX)23Hardware

Acceleration

forDLGPU

(Graphic

ProcessingUnit)Without

GPUDirectWithGPUDirectImage

source:

https://ASIC

designed

to

speed

up

DL

operations,

likeGoogle’s

TPU

(Tensor

Processing

Unit)HighperformanceLess

flexibleEconomical

only

at

large

scaleSpecialoptimizations

inhardwareFor

example:

reduced

precision,

matmul

operatorDesign

for

inference

is

different

from

thatfortrainingFor

example:

in

1st

generation

TPUs

fp-units

werereplacedbyint8-unitsHardware

Acceleration

forDLASIC

(Application

Specific

Integrated

Circuit)Image

source:

/nips17/assets/slides/dean-nips17.pdf24Hardware

Acceleration

forDLFPGA

(Field

Programmable

Gated

Arrays)Designed

to

bereconfigurableFlexibility

to

change

as

neural

networks

and

new

algorithms

evolveOffer

much

higher

Performance/Watt

than

GPUsCost

effective

and

excel

at

inferenceReprogramming

an

FPGA

is

not

easyLow

levellanguageAvailable

on

cloud

(EC2

F1instances)25Primarily

limited

to

inference-onlySpecialSoCdesign

with

reduced

die

spaceEnergy

efficiency

and

memory

efficiency

is

more

criticalSpecialoptimization

to

support

specific

tasks-onlyFor

example:

speech-only,

vision-onlyExamples:

Apple’s

Neural

Engine,

Huawei’s

NPUHardware

Acceleration

for

On-device

AI26Software

FrameworksFrontendAbstracts

the

mathematical

and

algorithm

implementationdetails

of

Neural

NetworksProvides

a

high

level

building

blocks

API

to

defineneuralnetwork

models

over

multiple

backendsA

high

level

language

libraryBackendHideshardware-specific

programming

APIs

from

userOptimizes

and

parallelizes

the

training

and

inference

processto

work

efficiently

on

the

hardwareMakes

it

easier

to

preprocess

and

prepare

data

for

trainingSupports

multi-GPU,

multi-node

execution27Dataset&Data

FlowUsing

Tensorflow

as

referenceDataset

Transformation

–

ImageNet

ExampleRawdata

vs

TFRecordsRaw

data

is

converted

into

packed

binary

format

for

training

called

TFRecord

(One

time

step)1.2

M

image

files

are

converted

into

1024

TFRecords

with

each

TFRecord

100s

of

MB

in

size1624761281802322843363884404925445966487007528048569089601,0125,120FREQUENCYSIZE

(IN

KB)Raw

ImageNet

Data10000

900080007000600050004000300020001000002920406080100120126127128129130131132133134135136137138139140141142143144145146147148149150151152153154156158161FREQUENCYSIZE(INMB)TFRecord

ImageNetDataTensorFlow

Data

PipelineIO:

Read

data

from

persistent

storagePrepare:

Use

CPU

cores

to

parse

andpreprocess

dataPreprocessing

includes

Shuffling,

datatransformations,

batching

etc.Train:

Load

the

transformed

data

ontothe

accelerator

devices

(GPUs,

TPUs)and

execute

the

DL

modelRead

IOPrepareTrainStorageNetworkCPU/RAMGPUPCIe/NVLinkGPUPCIe/NVLinkTFRecordsHost30WithoutpipeliningCompute

PipeliningWith

Pipelining

(using

prefetch

API)Image

source:

https:///guide/31Parallelize

IO

and

Prepare

PhaseParallelizeprepareParallelize

IOImage

source:

https:///guide/3