+ All Categories
Home > Documents > Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average...

Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average...

Date post: 08-Jul-2020
Category:
Upload: others
View: 0 times
Download: 0 times
Share this document with a friend
31
A A 2001251620/474996_2/474996/463867 Cover-Full Image-ArtworkNONO Neeta Madahar, Falling 1 (detail), 2005, Lightjet print mounted on aluminum, 48 x 48 inches (122 x 122 cm) Machine Learning Wellington Management Company LLP For institutional use only. Not intended for reproduction or use with the public. Any views expressed herein are those of the author(s), are based on available information, and are subject to change without notice. Individual portfolio management teams may hold different views and may make different investment decisions for different clients. The material and/or its contents are current as of the most recent quarter end, unless otherwise noted. Certain data provided is that of a third party. While data is believed to be reliable, no assurance is being provided as to its accuracy or completeness. 05 April 2018 Brian Barbetta Global Industry Analyst Wellington Management Company LLP
Transcript
Page 1: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

A

A 2001251620/474996_2/474996/463867

Cover-Full Image-ArtworkNONO

Neeta Madahar, Falling 1 (detail), 2005, Lightjet print mounted on aluminum, 48 x 48 inches (122 x 122 cm)

Machine Learning

Wellington Management Company LLP

For institutional use only. Not intended for reproduction or use with the public. Any views expressed herein are those of the author(s), are based on available information, and are subject to change without notice. Individual portfolio management teams may hold different views and may make different investment decisions for different clients. The material and/or its contents are current as of the most recent quarter end, unless otherwise noted. Certain data provided is that of a third party. While data is believed to be reliable, no assurance is being provided as to its accuracy or completeness.

05 April 2018Brian Barbetta Global Industry AnalystWellington Management Company LLP

Page 2: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

Image courtesy of Jeff Dean and Andrew Nq, www.googleblog.blogspot.com

A

A5 April 2018

The “cat” that changed the worldTitle Line 2

2001255852/474996_1/474996/462208

A1 2

Page 3: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001250470/474996_1/474996/463867A

A

Machine LearningTitle Line 2

What is it?

Why now?

Applications

Investment implications

A

A5 April 2018 A1 3

Page 4: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157467/474996_0/463867/463867

A

A5 April 2018

What are computers good at?Title Line 2

Computer programs must be based on logic and math

Computers are designed to evaluate equations

Computer is better than people

7 + 7 = 1 4

1 5.7 / 1 43 = 0.1 09637

7 > 4 = True

1 00 < -5 = False

A1 4

Page 5: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157199/474996_1/463867/463867

A

A5 April 2018

What are computers good at?Title Line 2

Computer programs must be based on logic and math

Computers are designed to evaluate equations

What is this?

Computer is better than people

Computer is not better than people

7 + 7 = 1 4

1 5.7 / 1 43 = 0.1 09637

7 > 4 = True

1 00 < -5 = False

A1 5

Page 6: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157489/474996_1/463867/463867

A

A5 April 2018

It’s a fish!Title Line 2

Could we write rules to determine if this is a fish?

A1 6

Page 7: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157509/463867_4/463867/463867

A

A5 April 2018

Fish?Title Line 2

A1 7

Page 8: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157518/463867_3/463867/463867

A

A5 April 2018

Fish?Title Line 2

A1 8

Page 9: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001255849/474996_1/474996/463867A

A

What if we teach computers to learn?Title Line 2

Programming a computer to learn is easier than writing rules to describe the world

But…

Requires enormous amounts of data to train on

Requires an enormous amount of compute capacity

A

A5 April 2018 A1 9

Page 10: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

ImageNet is an image database organized according to the WordNet hierarchy in which each node of the hierarchy is depicted by hundreds and thousands of images | Source: http://www.image-net.org/about-stats

A

A5 April 2018

ImageNetTitle Line 2

2001157559/463867_6/463867/463867

Summary and statistics (updated 30 April 201 0)

Overall•Totalnumberofimages:14,197,122•Over40,000peopleinvolvedintaggingimages

Sheep Dog

A1 10

Page 11: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

1As of November 2016

2001166981/474996_0/465160/463867

| The examples shown are presented for illustrative purposes only and are not to be viewed as representative of actual holdings. It should not be assumed that an investment in these (or similar examples) have been or will be profitable. Actual holdings will vary for each client and there is no guar-antee that a particular client’s account will hold the examples presented. | Source: ImageNet, Large Scale Visual recognition Challenge results 2015

A

A5 April 2018

ImageNetTitle Line 2

2010 2011 2012 2013 2014 201570

80

90

100

Human

74%

96%

Traditional CV

Deep learning

Rules-based learning Artificial intelligence

Traditional Computer Vision (CV) model•Humancoded•Sequential,linear•Timeintensive

Deep learning•Neuralnetwork-based•Parallelprocessingmade(newGPUs/hugedatasets)•Superfastandself-improving

Example: Object recognition competition resultsDeep learning achieves “superhuman” results (table charts accuracy rates, %)1

A1 11

Page 12: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157571/463867_3/463867/463867

A

A5 April 2018

How does it work?Title Line 2

Cat

A1 12

Page 13: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157572/463867_1/463867/463867

A

A5 April 2018

Why now?Title Line 2

A1 13

Page 14: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

Source: IDC‘s Digital Universe Study, sponsored by EMC, December 2012

A

A5 April 2018

DataTitle Line 2

The digital universe: 50-fold growth in data from the beginning of 201 0 to the end of 2020 (Exabytes)

2001157948/474996_1/463867/463867

2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 20200

10,000

20,000

30,000

40,000

Year

(Exabytes)

A1 14

Page 15: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

Sources: Ray Kurzweil, DFS (https://en.wikipedia.org/wiki/Moore%27s_law)

2001157950/463867_1/463867/462208

A

A5 April 2018

ComputeTitle Line 2

1 20 years of Moore’s Law

A1 15

Page 16: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157951/463867_4/463867/463867

A

A5 April 2018

ProofTitle Line 2

A1 16

Page 17: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

1Each point is the average score achieved per episode after the agent is run with e-greedy policy (e50.05) for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the curve is the average of the action-value Q computed over the held-out set of states. Note that Q-values are scaled due to clipping of rewards (see Methods). | Source: https://storage.googleapis.com/deepmind-data/assets/papers/DeepMindNature14236Paper.pdf

A

A5 April 2018

AI system playing Space Invaders over timeTitle Line 2

Average action value (Q)1Average score per episode2

2001157953/463867_3/463867/463867

0 200 400 600 800

1,000 1,200 1,400 1,600 1,800 2,000 2,200

0 40 80 120 160 200Training epochs

0

1

2

3

4

5

6

7

8

9

10

0 40 80 120 160 200Training epochs

A1 17

Page 18: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

Best Linear Learner is the best result obtained by a linear function approximator on different types of hand designed features12. Contingency (SARSA) agent figures are the results obtained in ref. 15. Note the figures in the last column indicate the performance of deep Q-network (DQN) relative to the human games tester, expressed as a percentage, that is, 100 3(DQN score2random play score)/(human score2random play score). |  Sources: “Human-level control through deep reinforcement learning”, Volodymyr Mnih1*, Koray Kavukcuoglu1*, David Silver1*, Andrei A. Rusu1, Joel Veness1, Marc G. Bellemare1, Alex Graves1, Martin Riedmiller1, Andreas K. Fidjeland1, Georg Ostrovski1, Stig Petersen1, Charles Beattie1, Amir Sadik1, Ioannis Antonoglou1, Helen King1, Dharshan Kumaran1, Daan Wierstra1, Shane Legg1 & Demis, ©2015 Macmillan Publishers Limited.Hassabis1

A

A5 April 2018

More gamesTitle Line 2

Breakout

31 .8

Humans

DeepMind

401 .2

2001250471/474996_1/474996/463867

A1 18

Page 19: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

Nature Publishing Group’s permission must be acknowledged next to the figure, table or abstract in print. In electronic form, this acknowledgement must be visible at the same time as the figure/table/abstract, and must be hyperlinked to the journal’s homepage. | Reprinted by permission from Macmillan Publishers Ltd: Nature Publishing Group (Human-level control through deep reinforcement learning), copy-right 2015

A

A5 April 2018

More gamesTitle Line 2

2001157955/474996_1/463867/463867

A1 19

Page 20: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157956/463867_3/463867/463867

A

A5 April 2018

Machine Learning todayTitle Line 2

A1 20

Page 21: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

From 2017 Alibaba investor day

A

A5 April 2018

Online shoppingTitle Line 2

2001157957/463867_3/463867/462208

Search – A thousand people a thousand faces

Recommended for you, Brian

BRian’s amazon

A1 21

Page 22: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157958/463867_3/463867/462208A

A

Machine LearningTitle Line 2Media

Brian, because you watched…

A

A5 April 2018 A1 22

Page 23: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

Source: www.wordstream.com

2001157990/463867_3/463867/462208

A

A5 April 2018

Why do we care?Title Line 2

Lookalike audience

Your database Social media

Match

A1 23

Page 24: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

Source: https://deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/

A

A5 April 2018

Machine Learning – EnergyTitle Line 2

2001157991/474996_1/463867/463867

DeepMind AI reduced Google data center cooling bill by 40%

A1 24

Page 25: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001250475/474996_0/474996/463867

A

A5 April 2018

OK, Alexa

A1 25

Page 26: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001250474/474996_1/474996/474996A

A

Machine LearningTitle Line 2Retinal Scans

Diagnostics using just retinal scans and deep learning•Slightlybetterthantheleadingophthalmologistsatdiagnosingdiabeticretinopathy•Abletodetermineifapatientisasmokerover70%ofthetime•Couldpredictbloodpressurewithin11mmHg•Couldforecastcardiovasculareventwithinfiveyears70%ofthetime(comparabletomoreadvanceddiagnostics)

Sources: https://research.googleblog.com/2016/11/deep-learning-for-detection-of-diabetic.html; https://research.googleblog.com/2018/02/assessing-cardio-vascular-risk-factors.html

A

A5 April 2018 A1 26

Page 27: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001157986/463867_2/463867/463867

A

A5 April 2018

Why drive?Title Line 2

A1 27

Page 28: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001145211/462208_1/462208/462208

A

A5 April 2018

Machine LearningTitle Line 2What is needed to succeed?

Nozzle = algorithm Fuel = data

A1 28

Page 29: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

Source: http://webia.lip6.fr/~pajot/dataviz.html

A

A5 April 2018

Machine LearningTitle Line 2Contributers to machine learning

Bar plat of the number of accepted paper by institution

2001250609/474996_1/474996/474996

cs.ucla.edu

in.tum.de

inria.fruoguelph.ca

mail.huji.ac.il

Edimburgh University

ucla .edu

ens.frsocher.org

riken.jpkaist.ac.kr

umiacs.umd.edu

cs.utexas.edu

iit.itTsinghya

NVIDIA

mi.t.u-tokyo.ac.jp

uci.eduUPMC

uva.nlcs.huji.ac.il

Cornellttic.edu

Montreal University

seas.upenn.edu

University of Michigan

OpenAl

Columbia

Baidusalesforce.com

IntelPrinceton

NYUUSC

University of Illinois

Cambridge

Georgia Tech

Washington Univ.

MIT ETHToronto Univ.

IBM Oxford University

Microsoft

Facebook

CMUStanford

Berkeley

google.com

0

5

10

15

20

25

30

35

40

45

A1 29

Page 30: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2001255851/474996_1/474996/474996

A

A5 April 2018

Key takeawaysTitle Line 2

A1 30

Page 31: Machine Learning - Canadian Investment Revie...for 520 k frames on Space Invaders. | 2Average predicted action-value on a held-out set of states on Space Invaders. Each point on the

2000000298/462208_3/G1422/G1422

A

A5 April 2018

Important noticeTitle Line 2

©2017 Wellington Management. All rights reserved. | As of June 2017

Wellington Management Company llp (WMC) is an independently owned investment adviser registered with the US Securities and Exchange Commission (SEC). WMC is also a commodity trading advisor (CTA) registered with the US Commodity Futures Trading Commission. In certain circumstances, WMC provides commodity trading advice to clients in reliance on exemptions from CTA registration. In the US for ERISA clients, WMC is providing this material solely for sales and marketing purposes and not as an investment advic��� duciary under ERISA or the Internal Revenue Code. WMC ha����� nancial interest in o� ering its products and services and is not committing to provide impartial investment advice or give advice i����� duciary capacity in connection with those sales and marketing activities. WMC, along with its a� liates (collectively, Wellington Management), provides investment management and investment advisory services to institutions around the world. Located in Boston, Massachusetts, Wellington Management also has o� ces in Chicago, Illinois; Radnor, Pennsylvania; San Francisco, California; Beijing; Frankfurt; Hong Kong; London; Luxembourg; Singapore; Sydney; Tokyo; and Zurich. This material is prepared for, and authorized for internal use by, designated institutional and professional investors and their consultants or for such other use as may be authorized by Wellington Management. This material and/or its contents are current at the time of writing and may not be reproduced or distributed in whole or in part, for any purpose, without the express written consent of Wellington Management. This material is not intended to constitute investment advice or an o� er to sell, or the solicitation of an o� er to purchase shares or other securities. Investors should always obtain and read an up-to-date investment services description or prospectus before deciding whether to appoint an investment manager or to invest in a fund. Any views expressed herein are those of the author(s), are based on available information, and are subject to change without notice. Individual portfolio management teams may hold di� erent views and may make di� erent investment decisions for di� erent clients.

In Canada, this material is provided by Wellington Management Canada l lc , a US SEC-registered investment adviser also registered in the provinces of Alberta, British Columbia, Manitoba, New Brunswick, Newfoundland and Labrador, Nova Scotia, Ontario, Prince Edward Island, Quebec, and Saskatchewan in the categories of Portfolio Manager and Exempt Market Dealer. In the UK, this material is provided by Wellington Management International Limited (WMIL)����� rm authorized and regulated by the Financial Conduct Authority (FCA). This material is directed only at persons (Relevant Persons) who are classi� ed as eligible counterparties or professional clients under the rules of the FCA. This material must not be acted on or relied on by persons who are not Relevant Persons. Any investment or investment service to which this material relates is available only to Relevant Persons and will be engaged in only with Relevant Persons. In Germany, this material is provided by Wellington Management International Limited, Niederlassung Deutschland, the German branch of WMIL, which is authorized and regulated by the FCA and in respect of certain aspects of its activities by the Bundesanstalt für Finanzdienstleistungsaufsicht (BaFin). This material is directed only at persons (Relevant Persons) who are classi� ed as eligible counterparties or professional clients under the German Securities Trading Act. This material does not constitute investment advice, a solicitation to invest i��� nancial instruments o��� nancial analysis within the meaning of Section 34b of the German Securities Trading Act. It does not meet all legal requirements designed to guarantee the independenc������ nancial analyses and is not subject to any prohibition on dealing ahead of the publicatio������ nancial analyses. This material does not constitute a prospectus for the purposes of the German Capital Investment Code, the German Securities Sales Prospectus Act or the German Securities Prospectus Act. In Hong Kong, this material is provided to you by Wellington Management Hong Kong Limited (WM Hong Kong), a corporation licensed by the Securities and Futures Commission to conduct Type 1 ( dealing in securities), Type 2 (dealing in futures contracts), Type 4 (advising on securities), and Type 9 (asset management) regulated activities, on the basis that you are a Professional Investor as de� ned in the Securities and Futures Ordinance. By accepting this material you acknowledge and agree that this material is provided for your use only and that you will not distribute or otherwise make this material available to any person. In Singapore, this material is provided for your use only by Wellington Management Singapore Pte Ltd (WM Singapore) (Registration Number 201 41 5544E). WM Singapore is regulated by the Monetary Authority of Singapore under a Capital Markets Services Licence to conduct fund management activities and is an exemp��� nancial adviser. By accepting this material you represent that you are a non-retail investor and that you will not copy, distribute or otherwise make this material available to any person. In Australia, Wellington Management Australia Pty Ltd (WM Australia) (ABN1 9 1 67 091 0 90) has authorized the issue of this material for use solely by wholesale clients (as de� ned in the Corporations Act 2001 ). By accepting this material, you acknowledge and agree that this material is provided for your use only and that you will not distribute or otherwise make this material available to any person. Wellington Management Company llp is exempt from the requirement to hold an Australia��� nancial services licence (AFSL) under the Corporations Act 2001 i n respec������ nancial services. A registered investment adviser regulated by the SEC, among others, is exempt from the need to hold an AFSL fo��� nancial services provided to Australian wholesale clients on certain conditions. Financial services provided by Wellington Management Company llp are regulated by the SEC under the laws and regulatory requirements of the United States, which are di� erent from the laws applying in Australia. In Japan, Wellington Management Japan Pte Ltd (WM Japan) (Registration Number 1 99504987R) has been registered as a Financial Instruments Firm with registered number: Director General of Kanto Local Finance Bureau (Kin-Sho) Number 428. WM Japan is a member of the Japan Investment Advisers Association (JIAA) and the Investment Trusts Association, Japan (ITA). WMIL, WM Hong Kong, WM Japan, and WM Singapore are also registered as investment advisers with the SEC; however, they will comply with the substantive provisions of the US Investment Advisers Act only with respect to their US clients.

A1 31


Recommended