copyright © 2018, technology futures, inc. 1 · title: microsoft powerpoint - vanston - ai...


Lawrence Vanston, Ph.D.President,Technology Futures, Inc.

TFI 2018January 25-26, 2018

Marriott Courtyard DowntownAustin, Texas

Forecasting Artificial Intelligence

84 Waller St • Austin, Texas 78702 (512) 258-8898 • www.tfi.com

[email protected]


Excerpted from:


Outline

• Fundamental Driving Forces – Old and New• Adoption Forecast for the New

Transformative Technologies• AI is in the Middle of Everything• Performance Trend Basics• Performance Trends for AI• Implications


Fundamental Driving Forces - 1990s View

Physical Movement Telecommunications

Analog Digital Communications

Low Bandwidth High Bandwidth

Wireline Wireless

Electronic Optical

Circuit Switching Packet Switching

Mass Marketing Niche Marketing


New Fundamental Driving ForcesHumans Machines

People Communicating Things CommunicatingComputing AI /Cognitive Computing

Screen-based Augmented/Virtual RealityClient/Server Cloud ComputingSmall Data Big DataSlow Data Fast Data

These overlap & support each other!


Humans MachinesRobots

Self Driving Vehicles

Drones


Kognit – Cognitive Assistants for Dementia Patients

http://www.slideshare.net/diannepatricia/kognit-cognitive-assistants-for-dementia-patients


Adoption Forecast for the New Transformative Technologies


TFI Wireless Generations Forecast


TFI Wireless Connection Forecast

0

100

200

300

400

500

600

700

1990 1995 2000 2005 2010 2015 2020 2025

Mill

ions

of C

onne

ctio

ns

Year Source: Technology Futures, Inc.Historical Data Source:CTIA

Wireless Forecast 2017

Total

Standard

M2M


U.S. Broadband Lifecycles

Lifecycle (t) = Substitution (t) – Next Substitution (t)

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

1995 2000 2005 2010 2015 2020 2025

Perc

enta

ge o

f Hou

seho

lds

Year

1.5Mb/s

6 Mb/s24

Mb/s50

Mb/s

All BroadbandHouseholds

Source: Technology Futures, Inc.

2016Broadband Access

100Mb/s

Data Source: FCC. Data excludes mobile wireless broadband

300 Mb/s& Above


Ultra-HD Households (aka 4K)

UHDTV Data Source (Red Squares): Strategy Analytics

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

2000 2005 2010 2015 2020 2025 2030

Perc

enta

ge o

f Hou

seho

lds

YearSource: Technology Futures, Inc.

HDTVHouseholds

Ultra-HDTVHouseholds


HD

TV 216

HDTVAnalogous

Rate

Historical data sources: 2001-2004 Misc, 2005-2015 Leichtman Research

PublishedForecasts


Major Transitions of our Time

0%10%20%30%40%50%60%70%80%90%

100%

1990 2000 2010 2020 2030 2040 2050

Perc

ent S

ubst

itutio

n

Year Source: Technology Futures, Inc.

Forecasting Future

Broadband, Digital, Fiber, Wireless, IP

2017

Machines, AI, AR/VR, IoT


CleverProgramming

INPUTS OUTPUTS

CLOUD

AI is in the Middle of Everything

AI

AI

AIAI

AI

AI


Performance Trends Basics


1

10

100

1,000

10,000

100,000

1,000,000

1965 1970 1975 1980 1985 1990 1995 2000 2005 2010Year

Tran

sist

ors

in 1

000s

40048080

8086

286386™ Processor

486™ DX Processor

8008

Pentium® Processor Pentium® II Processor

Pentium® III ProcessorPentium® 4 Processor

Moore’s Law


Constant Percentage Rate of Advance (Exponential)• Most new technologies progress this way• Measure must reflect utility• Rates will continue if:

– The improvement is technically possible– Utility and demand continues– Basic approach remains the same

• If approach changes, look for– Discontinuities– Changes in improvement rate


Analog Modem & Broadband Speeds

28.814.4

2.4

0.3

1.2

9.6

56.0

100000

24000

6000

1500

0

1

10

100

1000

10000

100000

1000000

1980 1990 2000 2010 2020

Year

Nom

inal

Dat

a R

ate

(Mb/

s)

AnalogModems

Broadband

Performance Increases 4 times every 4 years

(42% annually)

Broadband Access 2003



Maximum Data Rates for Commercial Fiber Optic Systems

0.01

0.10

1.00

10.00

100.00

1,000.00

10,000.00

100,000.00

1,000,000.00

1980 1985 1990 1995 2000 2005 2010 2015 2020Year

Gig

abits

/Sec

ond

(Gb/

s)

Single Wavelength

DWDM


Fiber Data R

atesPre-2000 Trend

Assumes 150 Tb/s Limit; estimates vary

Post-bubble Blues

2011


Source: Raghunath Nambiar , Cisco, http://tech4b.blogspot.com/2012/04/fast-data-brief-analysis-on-trends.html

Average transactions-per-minute per processor


Performance Trends for AI


By Courtesy of Ray Kurzweil and Kurzweil Technologies, Inc. en:PPTExponentialGrowthof_Computing.jpg, CC BY 1.0, https://commons.wikimedia.org/w/index.php?curid=3324354

Note this isdouble exponential growth! -LKV

Kurzweil’sComputerPerformanceForecast

This isexponentialgrowth!


By Steve Jurvetson -https://www.flickr.com/photos/jurvetson/31409423572/, CC BY 2.0, https://commons.wikimedia.org/w/index.php?curid=55002144


Exponential Performance Trends

• Exponentialf(t) = ebt

• Multiple Exponentialsf(t) = eb1t eb2t = e(b1+b2 )t = ebt

Note that Multiple Exponential is Exponential• True Double Exponential

f(t) = eb = e(b )xx

Typical

Happens

NO!


Multiple vs Double Exponentials

1

10

100

1,000

10,000

100,000

1,000,000

10,000,000

100,000,000

1,000,000,000

0 5 10 15 20 25

Perf

orm

ance


Sample Trend

2017

Exp 2

Multiple Exponential

Exp 1 * Exp 2

DoubleExponential

Exp 1


Elon Musk on AI Progress and Threat at International Space Station Research and Development Conference, July 19, 2017

Full interview at https://www.youtube.com/watch?v=BqvBhhTtUm4

Clip from CNN Money/Tech


A Premise for Moving Forward

• Kurzweil means true double exponential, based on logic that better argues for multiple exponential.

• Musk means multiple exponential when he says “double exponential” (I think)

• Let’s assume that exponential is the most solid basis to explore the promise and impact of AI

• However, there’s a catch


Hard vs Easy Problems

• Easy Problems– Computations grow linearly or polynomially

with problem size n.– f(n) = a + bn– f(n) = a + b1n + b2n2 + b3n3 ... + bknk

• Hard Problems– Computations grow exponentially with problem

size n– f(n) = ecn


1 5 25 125

Decision Tree


Performance Trends for Problem Solving - Advanced• Exponential increase in processing speed, x(t)=eb1t

• Exponential increase in efficiency, y(t)=eb2t

• Combined performance, e(b1+b2)t = ebt

• Solution time = s(t) = f(n)/ ebt = e-bt f(n)• The maximum problem size solvable in S seconds

grows with n:Exponentially if f(n) is polynomial (Easy)Linearly if f(n) is exponential (Hard)


Maximum Problem Size in S seconds


Maximum Problem Size in S seconds (Example)

0102030405060708090

100

0 5 10 15 20 25

Max

Pro

blem

Siz

e in

S S

econ

ds


Sample Trend

Polynomial Difficulty

(Easy Problems)

2018

Exponential Difficulty (Hard Problems)

1000S =


Chess

https://www.eff.org/ai/metrics

Discontinuity Due to Hardware Improvement, Not Deep Learning


Go• AlphaGo performed notably better

than its predecessors (all Monte Carlo Tree Search programs, but without deep learning or GPUs.)

https://srconstantin.wordpress.com/2017/01/28/performance-trends-in-ai/

• It uses much more hardware and more data.• AlphaGo doesn’t represent that much of a surprise

given the improvements in hardware and data (and effort). – Miles Brundage

Miles Brundage (http://www.milesbrundage.com/blog-posts/alphago-and-ai-progress)


Prior AI Go Improvement

Found at: https://en.wikipedia.org/wiki/Monte_Carlo_tree_searchReferenced source for data: "Sensei's Library: KGSBotRatings". Retrieved 2012-05-03. https://senseis.xmp.net/?KGSBotRatings


AlphaGo Continued Progress

https://deepmind.com/blog/alphago-zero-learning-scratch/


Image Recognition


Copyright © 2018, Technology Futures, Inc. 40https://srconstantin.wordpress.com/2017/01/28/performance-trends-in-ai/(http://itl.nist.gov/iad/mig/publications/ASRhistory/index.html

SPEECH RECOG-NITION

201020001990

2011: context-dependent deep neural network hidden Markov models


Speech Recognition



Machine Translation



Natural Language Processing (SAT)(a) SAT Analogies (b) SAT Multiple Choice

(c) SAT Paraphrase ID

https://srconstantin.wordpress.com/2017/01/28/performance-trends-in-ai/

a. Analogies: No Deep Learningb. Mult Choice: Neural Nets in 2014c. Paraphrase ID: Top Performer was Matrix Factorization


Generating computer programs from specifications



Summary for AI Progress

• Wide Range of: – Problem difficulty– Solution status– Metrics– Progress rates and patterns– Distance to go

• Discontinuities and rate changes• Deep learning has varied impact on rate• Linear progress typical


Other Thoughts on AI Progress

• What are the characteristics of problems that decide the above?

• Do the problems become easier or harder as we move toward General Intelligence?

• How will AI systems play with each other?• How will AI interact with the outside world?• Does the argument that AI will “bootstrap”

itself hold water?


What does it mean for...

• Commerce and Industry?• Communications?• Mankind?


For Commerce and Industry

• Wide range of applications across industries• Ubiquitous• AI embedded in other transformative

technologies• Important across processes, products, and

services• Many opportunities, but highly disruptive• Huge employment implications


Implications for Communications

• Rapidly increasing performance requirements• Rapidly improving technologies• Widespread and massive investments• Frequent upgrades and rapid obsolescence• Changes in workforce


Implications for Mankind

• Tremendous opportunity for good and evil• Existential Threats – Real and Imagined• Challenges – political, economic, and social


http://aiimpacts.org/wp-content/uploads/2017/04/ESOPAI-value.png


It’s up to us.

copyright © 2018, technology futures, inc. 1 · title: microsoft powerpoint - vanston - ai...

Documents