December 2019

Machine Learning
Innovation
A Patent Analytics Report

Prepared by IP Australia
THIRD PARTY COPYRIGHT
IP Australia prepared the report which has made all reasonable efforts to:
• clearly label material where the copyright is owned by a third party
• ensure that the third party has consented to this material being presented in this
publication.
Permission may need to be obtained from third parties to re-use their material.

ATTRIBUTION
The CC BY licence is a standard form licence agreement that allows you to copy and
redistribute the material in any medium or format, as well as remix, transform, and
build upon the material, on the condition that you provide a link to the licence, you
indicate if changes were made, and you attribute the material.

DISCLAIMER
The information contained in this brief has been gathered from global intellectual
property (IP) databases and represents a snapshot of IP rights at a particular point
in time. It is provided for general information only and should not be relied upon for
the purposes of any particular matter. It is not a report on patentability or freedom to
operate and should not be relied upon for those purposes.
 Foreword
Twelve months ago, ACS released Artificial Intelligence - A Starter
Guide To The Future Of Business. We wanted to assist businesses to
position themselves for a future that optimises machine-readable data
patterns, automates processes wherever possible, and enables new
digital product development that can enhance Australia’s exports.
Yohan Ramasundara
President, ACS The use cases identified in the Starter Guide covered Agriculture (AI-
controlled field robots, crop and soil monitoring, predictive analytics);
Finance (algorithmic trading, Robo-advisors, deep data analytics);
Health care (diagnostics, augmented decision making, AI-managed
administration); Mining (mineral exploration, autonomous equipment,
automated sorting and optimised metallurgy, predictive maintenance);
Manufacturing (self-learning monitoring, predictive maintenance,
intelligent supply chain optimisation, AI-driven product design); Retail
(product recommendations, product identification and personalisation,
demand prediction, customer engagement); Communications and
Entertainment (automated content creation, personalised marketing
Andrew Johnson and advertising, automated telemarketing and polling); and Transport
Chief Executive (autonomous vehicles and driver assistance, engineer monitoring and
Officer, ACS predictive maintenance, safety and efficacy).

Since the time of the Starter Guide’s release, we have seen revenue
from the global AI market forecast to increase more than tenfold from
US $9.5 billion in 2018 to an estimated US$118 billion by 2025i. We
have witnessed 14 of the world’s most advanced economies announce
over AU$86 billion in focused AI programs and activitiesii.

As momentum builds in this fourth industrial revolution, ACS

wanted to move beyond the hype and undertake an evidenced-based
investigation into machine learning progress.

We considered there to be no better place to research than patents

– both as a guide to see where the world’s innovators believe there
are gaps in the market that can be commercialised through machine
learning technologies, as well as potentially indicate those industries
and use cases that appear untapped, thus providing innovators with
insight as to where early leader advantage may still be achieved.

Having worked with the Patent Analytics Hub of IP Australia previously

on blockchain patent trends, it was once again our privilege to now
explore machine learning and artificial intelligence.

i
Tractica (30 April 2019), Artificial Intelligence Software Market to Reach $118.6 Billion in Annual Worldwide Revenue by 2025
https://www.tractica.com/newsroom/press-releases/artificial-intelligence-software-market-to-reach-118-6-billion-in-annual-
worldwide-revenue-by-2025/
ii
Artificial Intelligence Roadmap, Australia’s artificial intelligence roadmap, developed by CSIROs Data61 for the Australian
Government.

MACHINE LEARNING INNOVATION - DECEMBER 2019 3

 Contents

02
Introduction 
Why patent data?
8
9
04
Core capabilities  16
The authors and
purpose of this report 9 Top applicants 19

01
Image and video analysis 20
Analytics and processing 22

03
Speech and text analysis 22
Control and automation 23
Summary  6 Signal analysis 24
Other applications 24
Patenting
trend overview  10
Patent filings 10
Patent destination 11
Applicant origin 12
Technology overview 15

4
 06
Foundation
technologies 36
Top applicants 37

08
Algorithms 39
Architecture 42

Conclusion 48

05 09
Appendix A: Definitions 50
Real world Patent, applications and

07
applications 26 publications 50
Telecommunication 28
Patent families 50
Environment and mining 29
Classification 50
Energy 29
Patent status determination 50
Health care 30
Australian
applicants 44 Country of origin 50
Predictive maintenance 30
Appendix B: Search
Industry 30 strategy 52
Transport and logistics 32 Data extraction and analysis 52
Security and military 32 Search strategy 52
Agriculture 34 Appendix C: Technology
Business and finance 34 analysis 54
Retail 35 Abbreviations 64

MACHINE LEARNING INNOVATION - DECEMBER 2019 5

01
SUMMARY

6
Summary KEY REPORT FINDINGS

Machine learning technologies 36,740 patent families

have been filed in
underpin artificial intelligence (AI).
machine learning
Machine learning algorithms are applied technologies since 2012.
to make predictions and conclusions when 25,319 originate
provided with data and allow computer systems from China.
to make intelligent decisions.1
Machine learning patent
Recent patent data shows machine learning
filings have grown 27%
technology is an area of outstanding growth
or more every year
and investment across the world. Since 2012,
since 2012.
patenting of AI and machine learning related
technologies has increased significantly, with
a four-fold global increase in patent filings
91% of AI and machine
over five years from 2012−16. Applicants are
learning related patent
commonly large corporations or universities.
families are in an
China is both the largest patent filing active state.
destination and the largest source of patent
family filings in machine learning, with the The State Grid
State Grid Corporation of China playing the Corporation of China is
largest role in this sector. the top global innovator,
The United States, South Korea and Japan with 1,018 active patent
are the other major players in this sector, families (88% are active).
with Australia ranking tenth in the world as a
patent filing destination.
China is the largest filing
In real world applications of machine learning, destination in the world,
the telecommunications sector had the most while Australia is 10th.
patent filings (17 per cent of patents filed), likely
reflecting the growth and net worth of this
sector in the global economy. This is mirrored
by development of core capabilities in image Australian innovators
and video analysis (36 per cent of patents filed). rank 17th globally, with
59 patent families.
Underlying these advances, the most
significant developments in foundational
technologies since 2012 were found in The telecommunications
supervised or semi-supervised learning sector leads real world
techniques (38 per cent of all patent filings). applications of machine
learning, with 17% of
This report outlines findings from analysis
patent applications.
of patents filed since 2012, including
trends, innovators, filing destinations and Image and video analysis
commercial players in this space. is the largest core
capability application,
with 36% of patent
applications.
CSIRO Data61, Artificial Intelligence and Machine
1

Learning, Focus Area (accessed 9 October 2019)

7
https://www.data61.csiro.au/en/Our-Work/AI-and-
Machine-Learning
MACHINE LEARNING INNOVATION - DECEMBER 2019
02
Introduction
AI is the simulation of human Revenue from the global AI
intelligence processes by market is forecast to increase
machines. It is the term more than tenfold from US
given to a collection of $9.5 billion in 2018 to an
interrelated technologies, estimated $118 billion by 2025.4
modern algorithms and forms Understanding how AI and
of artificial programs or machine learning technologies
systems capable of perception, are commercialised, identifying
cognition, decision making potential partners and
and implementation without determining what capability
explicit guidance from a human currently exists in Australia
being.2 Machine learning is a will be important in assisting
form of AI that enables a system the Australian ICT sector to
– or machine – to learn from leverage opportunities both
data using algorithms. These nationally and internationally as
algorithms refine the system this technology matures.
by processing new data. As a
result, the system is trained
to adapt to changing patterns
and associations in the data,
allowing intelligent predictions
and decisions to be made.3

2
Dawson, D. et al. (2019) Artificial Intelligence: Australia’s Ethics Framework (A Discussion
Paper). Data61 CSIRO, Australia
https://consult.industry.gov.au/strategic-policy/artificial-intelligence-ethics-framework/
supporting_documents/ArtificialIntelligenceethicsframeworkdiscussionpaper.pdf
3
SAS, Machine Learning Analysis Insight (accessed 9 October 2019)
https://www.sas.com/en_au/insights/analytics/machine-learning.html
4
Tractica (30 April 2019), Artificial Intelligence Software Market to Reach $118.6 Billion in
Annual Worldwide Revenue by 2025
https://www.tractica.com/newsroom/press-releases/artificial-intelligence-software-
market-to-reach-118-6-billion-in-annual-worldwide-revenue-by-2025/

8
Why patent data? Kingdom Intellectual Property
Office (UKIPO)6. These reports
professional association
and peak body representing
Patents can be used as show that AI-related inventions Australia’s ICT sector. It
indicators of innovative activity. are booming and shifting from aims to deliver authoritative
By extracting and analysing theoretical to commercial independent knowledge and
data associated with patent application, with the field of technology insights, building
documents, we can measure machine learning dominating relevant technology capacity
aspects of inventive activity this growth. and capability that catalyses
such as scope, intensity, Australian innovation, and speeds
collaboration and impact. The authors the adoption of technology
for the benefit of commerce,
These metrics can be developed
across technology sectors
and purpose government and society.
and by measures including of this report ACS is at the forefront
individuals (inventors),
This report provides a of education and policy
institutions (applicants),
specific analysis of emerging development in Australia’s ICT
countries and regions.
technologies in machine sector, contributing to building
Patents are granted for devices, learning relating to AI. capability in the development
substances, methods or of, and response to, emerging
IP Australia is dedicated to technologies. The 2019 ACS
processes that are new, inventive
building prosperity for Australia Australia’s Digital Pulse report
and useful, giving exclusive
and ensuring that Australians examines digital workforce
commercial rights in exchange
benefit from great ideas. Using trends and takes stock of
for full public disclosure of
patent data analytics to provide Australia’s digital economy.7
the invention. This means
evidence of innovation trends,
patents are a source of data on
we leverage our unique access In this patent analytics report we
innovation trends in science and
to IP data, knowledge and have analysed the development
technology. More information
expertise to deliver value to the and uses of technology relating
about the patent system is given
broader community. to machine learning.
in Appendix A: Definitions.
This report also provides
This report was prepared in insight into innovators in this
Technology trends across the
partnership with the Australian field – based on patents filed –
broad field of AI have been
Computer Society (ACS) as part with potential connections for
analysed using patent data
of IP Australia’s commitment research, development
in recent publications by the
to support Australian research and commercialisation.
World Intellectual Property
and innovation. ACS is the
Office (WIPO)5 and the United

5
World Intellectual Property Organization (2019), WIPO Technology Trends 2019, Artificial Intelligence. WIPO, Geneva
https://www.wipo.int/edocs/pubdocs/en/wipo_pub_1055.pdf
6
Intellectual Property Office (2019) Artificial Intelligence: A worldwide overview of AI patents and patenting by the UK AI sector. IPO,
Newport https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/808891/AI-a-worldwide-
overview-of-AI-patents.pdf
7
Australian Computer Society (2019) ACS Australia’s Digital Pulse 2019,
https://www.acs.org.au/insightsandpublications/reports-publications/digital-pulse-2019.html

MACHINE LEARNING INNOVATION - DECEMBER 2019 9

03 Patenting trend
overview
As a basis for this analysis, we searched the PATSTAT database (2018
Autumn edition) for inventions relating to machine learning. The
search8 found 36,740 unique DOCDB patent families (see Appendix A:
Definitions) relating to machine learning filed worldwide since 2012.
This year range allows a focus on the most recent patent filings and
includes complete patent data over five years.

Patent filings Global patenting activity relating

to machine learning has
or withdrawn may be reflective of
both recent patenting activity in
Analysing patent family filings increased dramatically from this technology area, and a high
across time indicates growth or 2012–16, with more than a four- commercial interest in AI and
decline in innovation. Additional fold increase. The dip in patent machine learning.
in-depth analysis can help filings from 2017–18 reflects
understand underlying factors and incomplete data rather than a
their correlation to these trends. declining trend.

Figure 1 shows the number of Of the 36,740 patent families

patent families by their earliest filed since 2012, 33,505 (91 per
priority year and patent status. cent) are in an active state with
More details on patent status patent protection being sought or
are given in Appendix A: Patent in force. The low levels of patent
status determination. families that have lapsed, expired

FIGURE 1 PATENT FAMILIES RELATING TO MACHINE LEARNING BY EARLIEST PRIORITY YEAR AND PATENT STATUS
(Note: data for 2017–18 is incomplete due to a lag in publication of patents)
Source: PATSTAT 2018 Autumn Edition

10,547
9,950
Number of patent families

6,030

4,080
3,204
2,410

519
Being sought or in force

2012 2013 2014 2015 2016 2017 2018 Lapsed, expired or withdrawn

8
The search used a combination of keywords, International Patent Classification (IPC) symbols and Cooperative Patent Classification

10
(CPC) symbols. Detail of the search methodology is provided in Appendix B: Search strategy
Patent destination Figure 2 shows the number
of patent families filed in
The United States is second
with 22 per cent of the total,
Applicants must file patent different jurisdictions. Patent representing less than a third of
applications in each country Cooperation Treaty (PCT) patent filings in China. In third,
or patent jurisdiction where applications have been excluded fourth and fifth place are South
they wish to have enforceable from this figure because these Korea, Japan and Europe with
patent protection. Reasons do not represent an enforceable only 4–5 per cent of total patent
for patent protection include right in any jurisdiction. families filed in each.
the country being a target for
China is the largest patent Australia ranks tenth in patent
commercialisation, further
destination, with 73 per cent of destinations for machine
research, or manufacturing.
total machine learning related learning related patents, with 295
patent families filed here. patent families filed in Australia.

FIGURE 2: PATENT FAMILIES RELATING TO MACHINE LEARNING BY PATENT DESTINATION

(Note: only countries with more than 100 patent family filings are listed)
Source: PATSTAT 2018 Autumn Edition

CHINA 26,758 UNITED KINGDOM 429

UNITED STATES 8,064 TAIWAN 384
SOUTH KOREA 1,823 CANADA 348
JAPAN 1,693 AUSTRALIA 295
EUROPE 1,611 RUSSIA 255
GERMANY 472 BRAZIL 107

MACHINE LEARNING INNOVATION - DECEMBER 2019 11

11
 Applicant origin Country information was
derived from applicant
Full details of this approach are
given in Appendix A: Definitions.
Analysis of patent applicant address and applicant
In the case of applicants who
origin provides insights into name harmonisation, where
are resident in China, coverage
countries with investment or applicant address information
of country information in
interest in a specific area of was available. If a patent had
PATSTAT is relatively poor.
innovation. Figure 3 shows the applicants with addresses
These applications confirmed
top 10 countries where patent from more than one country,
the validity of the approach
families relating to machine representing international
we used to derived applicant
learning originate. The outer collaboration, the patent was
country information, because
bars represent the total number counted separately for each
most of these patent families
of families from a given country country of applicant origin.
were also solely filed in China.
of origin and the inner bars Where applicant country code
represent the number of families information was not available This method of applicant country
from that location that have at (i.e. null data that would assignation may account for
least one member filed with an result in the patent application the relatively larger number
authority other than the home not being assigned to any of Chinese applications found
application authority (for example country), the country of the in the present dataset than in
a Chinese applicant filing in the application authority of the the recent WIPO9 and UKIPO10
United States or filing a PCT first application filed in the patent analytics reports on AI.
application with WIPO). family was used as a proxy.

FIGURE 3 PATENT FAMILIES RELATING TO MACHINE LEARNING BY APPLICANT ORIGIN

Source: PATSTAT 2018 Autumn Edition

CHINA 25,319 907

UNITED STATES 6,590 3,127
SOUTH KOREA 1,630 573
JAPAN 1,503 997
GERMANY 427 372
UNITED KINGDOM 338 248
TAIWAN 327 218
CANADA 239 221
Number of families
SWITZERLAND 224 224
INDIA 144 130 Number of international families

9
WIPO (2019) WIPO Technology Trends 2019, Artificial Intelligence. WIPO, Geneva,
https://www.wipo.int/edocs/pubdocs/en/wipo_pub_1055.pdf
10
IPO (2019) Artificial Intelligence: A worldwide overview of AI patents and patenting by the UK AI sector. IPO, Newport
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/808891/AI-a-worldwide-

12
overview-of-AI-patents.pdf
China is the largest source of AI commercialisation. In 2015, the role of AI alongside other
patent filings, mirroring the result the United States government emerging technologies including
for patent destinations, with 69 invested approximately US $1.1 the Internet of Things, cloud
per cent (25,319) of patent family billion in unclassified research computing, big data analytics,
filings having a Chinese applicant. and development in AI-related and mobile technologies.15 In
Only 4 per cent of these families technologies. In 2016, the 2016, South Korea also released
from Chinese applicants have National Science and Technology a plan to invest US $2 billion by
international family members. Council of the United States 2022 to strengthen its AI research
chartered a subcommittee on and development capability and
The number of Chinese-
Machine Learning and Artificial reach the ‘global top four’.16 The
originating patents in machine
Intelligence to help coordinate program included establishing
learning may reflect Chinese
federal activities in AI.13 The latest six new AI research institutes to
government research and
American AI Initiative,14 issued by ramp up AI innovation.
development initiatives. These
the United States President this
have been particularly strong in Japanese government initiatives
year, is expected to drive further
the field of AI, with a government in AI research include the fifth
growth in AI and machine learning
venture capital investment of US Science and Technology Basic
related inventions.
$30 billion and state and public- Plan 2016–20. This sets out a goal
private investments totalling South Korea and Japan are in third of Japan leading in this sector
tens of billions more. 11 and fourth place respectively, each and transforming society with
with 4 per cent of global patent new information technologies
As well as this, the State Council
families. These originate from and systems.17 Japan also aims
of China released the ‘New
at least one applicant of either to train up to 250,000 AI experts a
Generation Artificial Intelligence
country (1,630 and 1,503 patent year to drive AI innovation.18
Development Plan’ in July 2017,
filings, respectively). Thirty-five
which aims to build a domestic Germany ranks fifth globally for
per cent of patents originating in
AI industry worth US $150 billion patent applicant origin with one
South Korea and 66 per cent of
and to become the leading AI per cent (427 patent applications)
patents originating in Japan have
power by 2030.12 of global patents originating
international family members filed.
The United States comes in second here, of which 87 per cent have
with 18 per cent (6,590) of patent South Korea is one of the biggest international family members.
families originating from at least government investors in AI
Germany has one of the largest
one American applicant. Forty- per capita. The South Korean
AI research centres in the world,
seven per cent of patent families Ministry of Science, ICT and
the German Research Center for
from United States applicants have Future Planning defined an
Artificial Intelligence (DFKI), with
international family members. Artificial Intelligence Information
an annual budget of around EUR
Industry Development Strategy
The United States has a range of 45 million.19
in 2016. The strategy considers
government measures supporting

11
Davenport, T. (27 February 2019) China is catching up to the US on artificial intelligence research, The Conversation
https://theconversation.com/china-is-catching-up-to-the-us-on-artificial-intelligence-research-112119
12
Future of Life Institute, AI Policy – China (accessed 9 October 2019) https://futureoflife.org/ai-policy-china/
13
National Science and Technology Council, Networking and Information Technology Research and Development Subcommittee
(2016) The national artificial intelligence research and development strategic plan,
https://www.nitrd.gov/PUBS/national_ai_rd_strategic_plan.pdf
14
Future of Life Institute, AI Policy – United States (accessed 9 October 2019)
https://futureoflife.org/ai-policy-united-states/?cn-reloaded=1
15
Future of Life Institute, AI Policy – South Korea (accessed 9 October 2019) https://futureoflife.org/ai-policy-south-korea/
16
Peng, T (17 May 2018) South Korea Aims High on AI, Pumps $2 Billion Into R&D, Medium, Synced Review,
https://medium.com/syncedreview/south-korea-aims-high-on-ai-pumps-2-billion-into-r-d-de8e5c0c8ac5
17
Harris, P (June 2017) Research brief: developments in artificial intelligence (AI) in Japan and implications for Australia, Australian
Government, Department of Education and Training, https://internationaleducation.gov.au/International-network/japan/
PolicyUpdates-Japan/Documents/AI%20in%20Japan%20research%20brief_07-2017.pdf
18
Yamashita, M (30 March 2019) Japan aims to produce 250,000 AI experts a year, Nikkei Asian Review,
https://asia.nikkei.com/Economy/Japan-aims-to-produce-250-000-AI-experts-a-year
19
German Research Center for Artificial Intelligence (accessed 9 October 2019)
https://www.dfki.de/en/web/

MACHINE LEARNING INNOVATION - DECEMBER 2019 13

 In 2018, the German applicant, and 68 per cent of with technological development,
government adopted a these having an international rather than take the lead.22 The
national AI strategy and family member. More details on Australian Government has
invested EUR 3 billion toward Australian applications are given made a funding commitment of
AI research and development, in the final section of this reports, AU $29.9 million over four years
with the aim of developing on Australian applicants. announced in the 2018–19 Federal
world leading AI capability. 20 Budget to promote AI research
A recent survey by Deloitte, State
and commercialisation. This is
of AI in the Enterprise, shows
Germany may also benefit from committed to initiatives including
that, although most Australian
the European Union’s investment the Cooperative Research Centres
organisations are using AI service
of EUR 1.5 billion in AI research.21 Program, scholarships, education
technologies, Australia has a
programs and developing a
Australia ranks seventeenth perceived skills and capability
national AI framework.23
globally, with 59 patent families gap in the AI sector, with
having at least one Australian Australians aiming to keep up

Patenting incentives in China

China is a world leader in not appear to have stimulated largest international filer to
filing patent applications large numbers of patents to be 2017.26 This pattern of Chinese
across all technologies. This filed in other jurisdictions for patent filing by Chinese
is partly driven by government patent protection.25 applicants could also be driven
policies to increase patent by further Chinese government
The Chinese patent office
filings and enable research patent initiatives, ranging from
receives by far the largest
commercialisation. Initiatives subsidies for local filings offered
number of patent filings of any
include ‘Measures for the by provincial governments,
office in the world, but - as is also
Administration of Special Funds including reimbursement of fees
reflected in the data presented
for Subsidising Foreign Patent and postponing of fees in case of
here - more than 95 per cent
Applications’, issued by the financial difficulties, to tax breaks
of all Chinese-filed patents are
Chinese Ministry of Finance for companies with IP holdings,
seeking protection only within
in 2012. This helps domestic cash bonuses and funds for use
China. This means the their
applicants enter the international in commercialising patented
impact outside Chinese borders
market by funding PCT and Paris inventions,27 and even reduced
is relatively limited compared
Convention patent applications.24 prison sentences for criminals
with that of jurisdictions such
This has pushed China into the who have granted patents.28
as the United States, which
second position globally in PCT
remained substantially the
filings in recent years, but does

20
Made in Germany, The national AI Strategy of the Federal Government (accessed 9 October 2019) https://www.ki-strategie-
deutschland.de/home.html
21
Dutton, T (June 29 2018) An Overview of National AI Strategies, Medium,
https://medium.com/politics-ai/an-overview-of-national-ai-strategies-2a70ec6edfd
22
Jackson, P and Marshall, A (May 2019) Future in the balance? How countries are pursuing an AI advantage, Deloitte, https://www2.
deloitte.com/au/en/pages/technology/articles/how-countries-are-pursuing-ai-advantage.html
23
Budget Review 2018–19, Parliament of Australia, https://parlinfo.aph.gov.au/parlInfo/search/display/display.
w3p;query=Id%3A%22library%2Fprspub%2F5982057%22
24
Notice on Printing and Distributing the Measures for the Administration of Funding for Foreign Patent Special Funds
http://www.gov.cn/zwgk/2012-05/31/content_2149501.htm
25
China Drives International Patent Applications to Record Heights; Demand Rising for Trademark and Industrial Design Protection
https://www.wipo.int/pressroom/en/articles/2018/article_0002.html
26
WIPO Stats on Patent Application Filings Shows China Continuing to Lead the World
https://www.ipwatchdog.com/2017/12/12/wipo-stats-patent-application-filings-china-lead/id=90855/
27
How patent subsidies boost R&D
https://www.managingip.com/Article/3538144/How-patent-subsidies-boost-R-D.html?ArticleId=3538144
28
How get out jail early in China: buy an inventor’s idea and patent it

14
https://www.scmp.com/news/china/article/1681850/how-get-out-jail-early-china-buy-inventors-idea-and-patent-it
Technology overview
Analysing the technical content of patents can provide insights into specific technologies. These insights
highlight areas of commercial interest and can be used to inform research and commercialisation strategies.
In machine learning, we have considered these developments in terms of four intertwined functional layers.
While each of these layers plays an important role, they are not mutually exclusive, which is demonstrated in
the examples of this report.

FUNCTIONAL LAYER
Broad functions that use foundational algorithms to
Applications Core capabilities
provide general outcomes
Uses of core capability functions to provide solutions for
Real world applications
real world problems

Foundational Technologies Algorithms Basic algorithms and techniques

Hardware and systems that allow new technologies to

Architecture
work

We start by broadly categorising machine learning applications, A total of 94 per cent (34,376) of
the field into either applications machine learning foundational these patent families discuss
or foundational technologies to technologies, or both. Figure 4 machine learning applications.
give an idea of investment and the shows the focus of the 36,740 These applications include the
degree of focus on applications of patent families, with 60 per cent core capabilities of image and
these technologies. We then focus referring to both a foundational video analysis or speech and
more specifically on examining technology and its application (see text analysis as well as the
real world applications and core Appendix C: Technology Analysis). application of these technologies
capabilities (used to facilitate real Altogether, 66 per cent (24,091) to industry. These core
world applications) and reviewing of these patent families relate capabilities and their real world
the state of play in foundational to foundational technologies. applications are discussed in
technologies (architecture and These include architecture detail in the following sections.
foundational algorithms). and algorithms, which we will
The patent families identified explore later in this report.
in this analysis relate to

FIGURE 4 TECHNOLOGY OVERVIEW OF PATENT FAMILIES RELATING TO MACHINE LEARNING

Source: PATSTAT 2018 Autumn Edition

12,649

21,727

Applications only

Foundational technologies only

Application and Foundational
technologies
2,364

MACHINE LEARNING INNOVATION - DECEMBER 2019 15

04
CORE CAPABILITIES

16
Core
capabilities
There are 34,376 patent Here we initially analyse core
families that describe machine capabilities, followed by real
learning applications where world applications.
machine learning is applied to
We have divided machine
solve problems. These patent
learning patent applications
applications can describe both
by their core capability
the core capability (27,956
according to patent technology
patent families) being applied
classification, in the following
e.g. image and video analysis,
hierarchal order. This hierarchy
and/or its real world application
allows a comparison between
(25,768 patent families) e.g.
broad sectors.
agriculture or healthcare.

HIERARCHY CORE CAPABILITY EXAMPLES OF APPLICATIONS

• Speech recognition
1 Speech and text analysis • Natural language processing
• Audio processing and pronunciation

• Computer vision
2 Image and video analysis • Object recognition and tracking
• Image matching and captioning

• Systems control
Control and automation
3 • Automation
processes
• Process optimisation

• Optical, electric, and magnetic signals

4 Signal analysis • Signal detection
• Use in diagnosis

• Data mining
Data analytics and
5 • Information retrieval
cognitive processing
• Forecasting

MACHINE LEARNING INNOVATION - DECEMBER 2019 17

 Figure 5 shows the number of patent families in each category by their grant status; granted (where
at least one family member is granted in any jurisdiction) or patent application (a patent family with
members seeking patent protection, or in which all members are lapsed/withdrawn without grant). The

FIGURE 5 PATENT FAMILIES RELATING TO MACHINE LEARNING CORE CAPABILITIES BY THEIR GRANT STATUS
Source: PATSTAT 2018 Autumn Edition

IMAGE/VIDEO ANALYSIS 2,970 10,352

ANALYTICS AND PROCESSING 1,290 4,301
SPEECH/TEXT ANALYSIS 1,125 2,756
CONTROL AND AUTOMATION 1,033 1,576 Granted patents

SIGNAL ANALYSIS 913 1,640 Patent applications

image and video analysis and filtering, and information control, automation and
sub-category captures retrieval and forecasting. This optimisation. There are 2,609
patent applications relating is the second largest category patent families in this category,
to computer vision, image with a total of 5,591 patent of which 40 per cent have a
processing, object recognition, families, of which 23 per cent granted family member.
object tracking, image have been granted in at least
Signal analysis includes optical
matching and captioning. This one jurisdiction.
(light), electric, electro-chemical
sub-category accounts for 48
Speech and text analysis and magnetic signals, their
per cent of patent families
includes natural language detection and their use in
describing machine learning
processing (NLP) and diagnosis. There are 2,553 patent
core capabilities, totalling
understanding, semantics, and families, of which 36 per cent
13,322 patent families of which
pronunciation, and speech and have a granted family member.
22 per cent have a granted
audio processing. There are
family member. Next we will take a closer look
3,881 patent families in this sub-
at each of these categories and
The analytics and processing category, of which 29 per cent
their top applicants.
sub-category covers patent have a granted family member.
applications relating to data
The control and automation
mining, processing, conversion
sub-category include systems

18
 Top applicants processing and 49 per cent
for speech and text analysis
and processing and less than
20 per cent of patent families in
Chinese applicants have the originating from applicants any other application field.
most filings in each core in China. The United States
Figure 6 shows the top
capability¸ with 77 per cent of is second for speech and text
10 applicants who filed patent
patent applications for image analysis with 35 per cent of
applications relating to
and video analysis, 72 per cent patent families originating there.
machine learning applications
for control and automation, This is a focus of the United
by core capability.
71 per cent for signal analysis, States, which has 23 per cent of
66 per cent for analytics and total patent families in analytics

FIGURE 6 PATENT FAMILIES RELATING TO MACHINE LEARNING CORE CAPABILITIES BY APPLICANT

Source: PATSTAT 2018 Autumn Edition

SPEECH/
ANALYTICS & CONTROL & IMAGE/VIDEO SIGNAL
TEXT
PROCESSING AUTOMATION ANALYSIS ANALYSIS
ANALYSIS

STATE GRID 390 58 181 144 17

CHINESE ACADEMY OF SCIENCES 90 38 467 57 88

IBM 143 13 70 15 317

XIDIAN UNIVERSITY 19 4 414 17 8

ZHEJIANG UNIVERSITY 68 77 143 48 22

MICROSOFT 113 1 72 7 127

UESTC 27 16 220 30 25

TIANJIN UNIVERSITY 31 12 190 43 29

ALPHABET 62 8 86 7 138

TSINGHUA UNIVERSITY 63 24 119 28 45

The State Grid Corporation of analysis, and ranks in the top in analytics and processing and
China (State Grid), which is six for all other application image and video analysis.
also the leading applicant in sub-categories as shown in
Xidian University ranks
machine learning techniques, Figure 6. Notably, the Chinese
second in the image and
ranks top in patent filings Academy of Sciences also
video analysis sub-category.
relating to data analytics and ranks second in machine
This is clearly a focus of the
processing, signal analysis and learning foundational
university, comprising 83 per
field only applications. State technologies, discussed later
cent of their patent portfolio
Grid also has a rich patent in this report.
relating to machine learning
portfolio in the control and
IBM, Alphabet and Microsoft applications. They are also
automation, and image and
take the lead – first, second and the third top applicant in the
video analysis capabilities.
third, respectively – in patent category of machine learning
The Chinese Academy of applications relating to speech foundational technologies.
Sciences rank top in patent and text analysis. Each company
filings for image and video also has a strong patent portfolio

MACHINE LEARNING INNOVATION - DECEMBER 2019 19

 Image and process consisting of: firstly,
making a list of areas in the
to produce a vector of visual
features, which is fed into
video analysis image that may contain an a recurrent neural network
object of interest using a (RNN) to generate labels
Image and video analysis is
region proposal network (RPN) using language modelling and
the most rapidly growing core
trained to sub-categories such sentence generation. This
capability in this dataset, with a
as cars with frontal view or patent family has a member
seven-fold increase from 2012–
people standing; and secondly granted in the United States
17. This is largely due to recent
classifying and refining the and protection in being sought
developments in machine
list of possible objects with in Australia and China.
learning that vastly improve
an object detection network
the ability to process visual Canon, a Japanese multinational
(ODN). The ODN simultaneously
information. Video and image company specialising in
assigns both category and
analysis is employed across imaging and optical products,
sub-category classification
a broad range of sectors host a healthcare optics
and bounding boxes to objects.
including telecommunications research laboratory and an
Applications of this technology
and networks, medicine imaging systems research
focus on detection of cars,
and healthcare, predictive laboratory.31 Canon has filed
pedestrians and cyclists in
maintenance, electrical power 62 patent families relating to
outdoor scenes. This patent
and energy, and transport image and video analysis since
family has a member granted in
and logistics. 2012. Canon’s patent portfolio
the United States.
includes US2016155024, which
Examples of representative
Adobe Systems is an American describes classifying elements
patent applications relating
multinational computer in a video as foreground or
to image and video analysis
software company focusing background, also known as
are held by NEC Corporation,
on multimedia and creativity scene modelling. The technique
Adobe Systems and Canon.
software and digital marketing operates in two phases:
NEC, a Japanese multinational software.30 Adobe Systems first classifying all elements
technology company, operates has 63 patent families filed to be either foreground or
with a vision of continuously relating to image and video background; and second
seeking technologies to create analysis. The application zooming in on a foreground
social value. NEC focuses US2017200065 describes object and re-classifying
research and development techniques for automatically elements in the frame. This
on developing recognition AI, generating descriptions of second classification is more
analytics using AI, AI platforms, images. The method uses sensitive to foreground. The
communications, cybersecurity, weak supervision data (noisy method is both computationally
and data distribution data that is not closely curated inexpensive and able to handle
platforms.29 NEC has filed and may include errors, such similarities between elements (a
67 patent families relating as tags from social networks) so-called camouflage scenario),
to image and video analysis. to provide information to and may be useful for activity
Application WO2017079522 supplement concepts extracted detection, unusual object or
describes a method for automatically from the image. behaviour detection, and scene
detecting objects in an image A combined machine learning analysis. This patent family has
by using ‘subcategory-aware’ approach is taken, where a a member granted in the United
convolutional neural networks pre-trained CNN is used to States, with protection being
(CNNs). This is a two-step assign descriptive terms and sought in Australia.

29
NEC Technology Vision (accessed 9 October 2019) https://www.nec.com/en/global/rd/technologyvision/index.html
30
Adobe (accessed 9 October 2019) https://www.adobe.com/
31
Canon Research and Development (accessed 9 October 2019)

20
https://www.usa.canon.com/internet/portal/us/home/about/research-development/
CASE STUDY: GENERATIVE ADVERSARIAL NETWORKS (GANS)
Machine learning has generated technique allows the creation of uses a conditional GAN (one that
such remarkable developments photorealistic videos at a very incorporates classification to
over the past five years that low cost. These techniques have exert some control on the output)
it may be difficult to grasp also been used to replicate audio to extract features of the input
the breadth of their potential in a similar manner. GANs are image and a ‘U-Net’ (a form of
impact. The development of based on the idea of adversarial convolutional network, to decode
GANs is an example of the huge training,37 using two neural the features). The conditional GAN
advances in this field. networks; the generator and the allows for more explicit extraction
GANs are a class of deep discriminator. The generator of features from the input image.
neural networks introduced captures the data distribution AU2018100325 discloses a method
by Ian Goodfellow and other and creates samples that are for colouring black-and-white
researchers at the University of almost identical to the training films and photos in real time, also
Montreal in 2014.32 GANs have data set. The discriminator by using conditional GANs.
been identified as ‘the most tries to differentiate between Patent application
interesting idea of the decade’ original samples and those WO2018053340, filed by Twitter,39
by Yann LeCun, who oversees AI created by the generator. During describes a method of enhancing
research at Facebook.33 training, the generator tries to low resolution images to higher
improve its synthetic samples resolution equivalents (known as
GANs have a wide range of
and the discriminator tries to super-resolution reconstruction).
applications, mostly related to
catch the fakes, and so the This method uses a GAN trained
images and video capabilities.
process is called adversarial on corresponding low- and high-
They allow improvements in
training. Success occurs when resolution images, followed by
image resolution, reconstruction
samples from the generator are sub-pixel convolution to create
of scenes, three dimensional
convincing to the discriminator.38 high-resolution output. The ready
models, synthetic music and text,
and even applications in physics Patent applications relating to accessibility of GANs means that
and biology.34,35,36 One of the GANs were first filed in 2016, both positive and negative social
most well-known uses of GANs with only four patent families outcomes may result from their
in the public eye is generation that year. Patent filings surged use in creating realistic synthetic
of deepfakes. Deepfakes are to 77 in 2017, with wider content. To mitigate the resultant
synthesised images or videos application of GANs in the areas risks associated with GANs,
synthesised using machine of image and video processing, researchers are also working to
learning techniques called and speech and text analysis. find markers of the use of GANs
deep neural networks. This Two innovation patent applications in images.40 The power of this
morphs one person’s face to filed in Australia relate to type of technology is vast, and as
mimic someone else’s features GANs. AU2017101166 relates to with any emerging technology,
while preserving original facial synthesising images and describes regulation lags behind early
expressions, without the need a method for transferring style adoption. It remains to be seen
for any video editing skills. This between images. This method what impacts GANs will have in
commerce, science and society.

32
Goodfellow, I.J. et.al., “Generative Adversarial Nets”, NIPS, 2014 https://arxiv.org/pdf/1406.2661.pdf
33
Yann LeCun Quora Session Overview (August 2016) https://www.kdnuggets.com/2016/08/yann-lecun-quora-session.html
34
Hui, J (23 June 2018) GAN — Some cool applications of GANs, Medium,
https://medium.com/@jonathan_hui/gan-some-cool-applications-of-gans-4c9ecca35900
35
Mustafa et al., CosmoGAN: creating high-fidelity weak lensing convergence maps using Generative Adversarial Networks,
Computational Astrophysics and Cosmology 6 (1) 2019
https://comp-astrophys-cosmol.springeropen.com/articles/10.1186/s40668-019-0029-9
36
Zhavoronkov et al., Deep learning enables rapid identification of potent DDR1 kinase inhibitors, Nature Biotechnology, 37, 1038-
1040, 2019 https://www.nature.com/articles/s41587-019-0224-x
37
Adversarial Machine Learning Tutorial (accessed 9 October 2019) https://aaai18adversarial.github.io/
38
Kang, N (May 2017) Deepfakes: The Good, The Bad and the Ugly, Medium,
https://medium.com/twentybn/deepfake-the-good-the-bad-and-the-ugly-8b261ecf0f52
39
About Twitter Inc (accessed 9 October 2019) https://about.twitter.com/en_gb.html
40
Murphy, H (16 August 2019) Cyber security companies race to combat ‘deepfake’ technology, Financial Times,

21
https://www.ft.com/content/63cd4010-bfce-11e9-b350-db00d509634e

MACHINE LEARNING INNOVATION - DECEMBER 2019

 Analytics and Facebook is the world’s largest
social networking company
a customised news stream
to a user, and measuring the
processing with a focus on connecting amount of time that news items
individuals with content and are displayed on a user’s device.
Developments in data analytics
stories they care about most.41 These times, along with the
and business intelligence
Facebook’s researchers develop features present in each news
methods using machine
machine learning algorithms item, are used to train a classifier.
learning are evolving predictive
that rank feeds, ads and search This provides a feed ranked by
and prescriptive analytics
results, and create new text expected viewing time by the user.
to gain insights from big
understanding algorithms that This patent family has a member
data sources. Patent filings
keep spam and misleading granted in the United States.
in this area have increased
content at bay.42 Facebook has
almost five-fold from 2012–16.
filed 54 patent families relating Speech and text
Patent applications in this
core capability predominately
to analytics and processing. analysis
WO2015038297 discloses a
relate to electrical power and Development of machine
method for understanding a
energy, business and finance, learning techniques has enabled
user’s sentiment toward a
telecommunications and speech and character recognition
page or topic based on their
networks, retail and medicine to have human-like fidelity
interactions with it. Part of the
and healthcare sectors. and the ability to dynamically
method is the use of traditional
Examples of patent applications sentiment analysis techniques to adjust to accents, handwriting
in analytics and processing are identify individuals with a strong styles and variable layouts.
held by Microsoft, Facebook polarity toward particular content Speech and text analysis core
and Yahoo. to generate a list of ‘trusted capabilities include NLP and
Microsoft has 113 patent users’. The information from understanding, which has rapidly
families in this core capability. this set of users is then used progressed in the past five years.
WO2016164694 describes a to train a predictive machine AI tools can now beat humans in
system that presents the best learning module that can be used comprehension tests.44
information to a content feed. for predicting the sentiment of Our analysis shows a five-
The system uses predictive new users interacting with the fold increase in speech and
models (such as neural content. This patent family is text analysis related patent
networks or regression models) seeking protection in the United filings between 2012–16, from
to predict the trending likelihood States, Canada, Japan, Korea 270 patent families to 1210.
of digital entities (documents, and Australia. These patent applications are
people, locations etc) based Yahoo is an American web applied across broad sectors
on historical behaviour and services provider. Yahoo including telecommunications
context. In combination with researchers develop large scale and networks, medicine and
information about the user machine learning algorithms healthcare, business and
and their network, this allows relating to content ranking, finance, and retail.
presentation of the most powering ad selection, spam Examples of patent applications
relevant and trending digital detection and abuse prevention.43 in this sub-category are held
entities to that user. This patent Yahoo’s patent portfolio includes by IBM, Alphabet, Apple and
family is granted in the United 30 patent families on analytics NTT (Nippon Telegraph and
States and is seeking protection and processing. US2017308537 Telephone Corporation).
in China and Europe. describes a method for providing

41
Statista, Most famous social network sites 2019, by active users (accessed 9 October 2019)
https://www.statista.com/statistics/272014/global-social-networks-ranked-by-number-of-users/
42
Applying machine learning science to Facebook products (accessed 9 October 2019)
https://research.fb.com/category/machine-learning/
43
Yahoo Research Areas (accessed 9 October 2019) https://research.yahoo.com/research-areas
44
Chong, Z (16 January 2018) AI beats humans in Stanford reading comprehension test, CNET,

22
https://www.cnet.com/news/new-results-show-ai-is-as-good-as-reading-comprehension-as-we-are/
An IBM patent application devices and services, touching State Grid patent application
US2017323011 describes the lives of millions of users CN105262113 sets out a method
methods for answering every day.45 Apple has a total for controlling a photovoltaic
questions using a ‘persona- of 51 separate patent families power generation system using a
based’ NLP system. The filed in this sub-category. probabilistic fuzzy neural network
user selects the persona for WO2016003509 relates to a with an error backpropagation
the machine to adopt, and virtual assistant for facilitating learning algorithm. The system is
the machine considers the interactions with a television. The designed to rapidly adjust the in
personality attributes of that virtual assistant uses NLP and and out of phase power that the
persona in addition to any factual contextual information (such as generator puts into the grid, in
information in formulating its items displayed) to understand a order to robustly handle changes
response to queries. A member user’s intention and to accurately in grid conditions such as faults,
of this patent family has been play their desired media. This and to support grid voltage
granted in the United States. patent family includes members during failures. This patent family
One of Alphabet’s patent filed in the United States, Europe, has a member granted in China.
applications US2017186420, Australia, China and Taiwan, with
Ford Motor Company’s mission
originally filed by Google, patents granted in the United
is to design smart vehicles
describes a method for converting States, Australia and Taiwan.
for a smart world.46 The
audio input into a sequence of NTT is the fourth largest company focuses on using
phonemes (a set of unique sounds telecommunications company in AI and machine learning
that cover a particular language). the world, and is taking a leading across different sectors from
Acoustic features are extracted role in digital transformation connected car solutions to
at each time step, and these in the telecommunications autonomous vehicles. Their
features are processed using a sector. One example of a AI platform called CarStory
long-short-term memory (LSTM) patent application held by NTT analyses the inventory data
approach. LSTM neural networks in speech and text analysis is of used cars to gain insights
are a very successful embodiment WO2018062265 which describes into what makes each
of an RNN, adapted to understand the use of a predictive acoustic previously owned car unique.47
long-term dependencies in data, model to perform speech US2018099646 uses a neural
and are particularly useful for recognition with improved network to classify levels of
time-series data like speech efficiency and less training data. precipitation using a range
processing. The output of the
LSTM component of the system is Control and of vehicle sensors such as
vibration of the windshield
used to generate the probability
that the sound at any given time
automation and imaging (Figure 7). This
classification is used to
step is accurately represented by Patent filings in control and
control vehicle functions such
a particular phoneme, building automation core capabilities
as windshield-wiper speed
a map of received sound into tripled between 2012–16. This
and traction-control. This
language. This method can be analysis found applications
patent family includes family
applied to speech recognition were concentrated around
members actively seeking
along with pronunciation and electrical, power and energy,
protection in the United States,
language modelling. This industry and manufacturing,
China and Germany.
invention has been patented in the transport and logistics and
United States. predictive maintenance.

Apple are dedicated to developing Examples of patent applications of

machine learning technologies to this sub-category are held by State
incorporate intelligence into their Grid and Ford Motor Company.

45
Apple machine Learning Journal (accessed 9 October 2019) https://machinelearning.apple.com/
46
Ford-About us (accessed 9 October 2019) https://corporate.ford.com/company.html
47
Marr, B (17 May 2019)The Amazing Ways The Ford Motor Company Uses Artificial Intelligence And Machine Learning, Forbes,
https://www.forbes.com/sites/bernardmarr/2019/05/17/the-amazing-ways-the-ford-motor-company-uses-artificial-intelligence-

23
and-machine-learning/#313f79dee49a

MACHINE LEARNING INNOVATION - DECEMBER 2019

 10, 10a

14, 14c FIGURE 7 IMAGE FROM FORD

16, 16a MOTOR COMPANY’S PATENT
14, 14b 12 APPLICATION US2018099646
14, 14a

16, 16b

20
22
18
This patent application shows
24a
24b the use of machine learning
neural network in the core
30
capability of control and
automation. This links to the
10, 10b 26
real world use of a neural
network to control vehicle
32 28 functions such as windshield-
wiper speed and traction-
control based on vehicle
34 sensors such as windshield
vibration and imaging.

Signal analysis with software optimisation for

tasks.48 US2018206774 relates to
to correlate the outputs of these
various sensors to provide
Patent families relating to signal a wearable device for sufferers of clinically relevant information. This
analysis tripled from 2012–16, Parkinson’s disease. This device patent family includes a member
with 685 patent families filed in is able to identify a specific motor granted in the United States and
2016. Signal analysis is widely symptom of the disease called members actively seeking patent
used for detection and diagnosis Freezing of Gait (FoG) using a protection in Canada, China and
in various sectors including neural network, and uses heart South Korea.
medicine and healthcare, rate detection and inertial sensors
predictive maintenance, to identify FoG events. Patent Other
electrical power and energy, and protection is being sought in the applications
telecommunication and networks. United States.
The California Institute of There are 6,420 patent
Representative patent
Technology (Caltech) is a science applications, of which 35 per
applications from this core
and engineering institute with cent have a granted family
capability are held by Intel and the
research interests ranging member, that describe real
California Institute of Technology.
from quantum science and world applications of machine
Intel, the American computer learning, but which do not specify
company that provides engineering to bioinformatics,
human behaviour and economics, a core capability. While real world
computing, networking, data applications of machine learning
storage and communication and energy and sustainability.49
WO2018035257 describes an will be discussed in detail in the
solutions worldwide, uses Intel next section, analysis of this group
AI to handle their AI patenting. enhanced stethoscope with
a range of sensors (such as of patent families gives an idea of
Intel combines multi-purpose, movements in development from
customisable and application- stethoscopic and non-stethoscopic
audio, ultrasound, temperature foundational and core capabilities
specific solutions to provide the to real world applications.
best hardware for each job, along etc). Machine learning is used

48
Intel AI (accessed 9 October 2019) https://www.intel.ai/

24
49
Caltech at a Glance (accessed 9 October 2019) https://www.caltech.edu/about/at-a-glance
CASE STUDY: PRIVACY PROTECTION
Control of data privacy is in the development and use upload data and machine learning
increasingly important, as the of privacy preserving machine code, and the code can be run on
global economy is increasingly learning. Privacy preservation the data in a secure environment
driven by data. Tech corporations goals in machine learning are without either party having access
dominate market capitalisation approached in several ways, to both components. The output is
lists,50 and the major product of with leading fields including then provided to the third parties.
many of these companies is either differential privacy, secure To reduce the vulnerability of
data itself (including Alphabet and enclaves, federated learning and stored data, machine learning
Facebook) or strongly dependent homomorphic encryption. algorithms can also be run as
on the use of data, as well as the Differential privacy is a federated learning methods in
sale of data processing capability requirement that an algorithm a distributed fashion, reducing
(including Microsoft, Amazon, is not overly dependent on the the need for transferring data
Apple, Alibaba and Tencent). information from any individual into a centralised data store
The devices we use, the cities in producing its output. A giving data suppliers more
we inhabit, the hospitals we visit differentially private algorithm control over how their data
and the shops we patronise are makes it unclear whether any is used.55 WO2018057302 by
constantly collecting our data, individual’s data was used as Alphabet relates to efficient
both public and intimate. This the input leading to a particular communication techniques for
data is used in machine learning statistical output.52 Relevant patent performing distributed machine
applications to train advertising, applications include CN107368752 learning on unreliable networks,
prediction, language processing from the Beijing Technology and such as mobile networks with
and diagnostic machines, Business University,53 directed intermittent coverage.
amongst many others; but this toward protecting data used for Homomorphic encryption allows
data can also be highly sensitive. training deep learning models certain computations to be
Privacy considerations associated from attacks that use GANs to performed on the encrypted data
with data storage are highlighted recreate training datasets. This itself, without requiring access
by the use of machine learning is achieved by anonymising the to the raw data.56 CN107241182
techniques, which require a vast training data, running simulated from the University of Electronic
amount of data for training.51 attacks, and using the degree of Science and Technology of China
In addition, models trained on recreation of the input as feedback (UESTC) relates to a method of
private data can be attacked in for the anonymisation procedure. clustering vectors where the
order to recreate the private data. Machine learning models can vectors are homomorphically
Machine learning itself can be executed in secure enclaves, encrypted before being passed
also be used to, deliberately or which provide trusted and isolated to a machine for clustering
accidentally, create data links environments54 WO2017222902 calculations and then the
and inferences that would not from Microsoft discloses secure processed clusters are passed
be possible by humans. There is execution environments where back to the client for decryption.
therefore an increasing interest third parties can separately

50
Desjardins, J (21 June 2019) The Front of the Pack, How the make-up of the world’s largest companies has changed since 1999,
Visual Capitalist, https://www.visualcapitalist.com/a-visual-history-of-the-largest-companies-by-market-cap-1999-today/
51
Mancuso, J (11 January 2019) Privacy-Preserving Machine Learning 2018: A Year in Review, Medium,
https://medium.com/dropoutlabs/privacy-preserving-machine-learning-2018-a-year-in-review-b6345a95ae0f
52
Ketchell, M (19 March 2018) Explainer: what is differential privacy and how can it protect your data?, The Conversation,
http://theconversation.com/explainer-what-is-differential-privacy-and-how-can-it-protect-your-data-90686
53
Beijing Technology and Business University (accessed 9 October 2019) http://english.btbu.edu.cn/
54
The Oasis team (1 September 2018) Towards an Open-Source Secure Enclave, Medium, Oasis Labs,
https://medium.com/oasislabs/towards-an-open-source-secure-enclave-659ac27b871a
55
Tramel, E 2019 Federated Learning: Rewards & Challenges of Distributed Private ML,
https://www.infoq.com/presentations/federated-learning-distributed-ml/

25
56
Homomorphic Encryption (accessed 9 October 2019) https://brilliant.org/wiki/homomorphic-encryption/

MACHINE LEARNING INNOVATION - DECEMBER 2019

05 Real world
applications
The use of machine learning techniques in real world applications
demonstrates how comprehensively machine learning developments
affect our lives. Analysis of these applications gives us a glimpse of
what the future holds as this dramatic growth continues to develop.

Real world applications (25,768 patent families) of machine learning

build on the development of core capabilities and foundational
technologies. In this section, we will show how machine learning
is affecting specific industries. As part of the layered technical
development of machine learning, this section demonstrates the
potential and complexity of this field.

Here we have divided machine learning patent applications by their

real world application according to patent technology classification,
in the following hierarchal order. This hierarchy allows a
comparison between broad sectors, despite obvious cross-over
between different applications.

26
HIERARCHY REAL WORLD APPLICATION EXAMPLES OF APPLICATIONS

• Monitor equipment, structures or vehicles

1 Predictive maintenance • Predict when maintenance is required
• Faults detection

• Managing mobile and computer networks

2 Telecommunication • Internet and data security
• Automated telemarketing and polling

• Field robots
3 Agriculture • Crop and soil monitoring
• Predictive analytics

• Generating energy
4 Energy • Automated management of grids
• Renewable energy

• Mineral exploration
5 Environment and mining • Automated sorting and optimised metallurgy
• Earth observation

• Biometric scanning
6 Security and military • Fraud detection
• Reading sentiment in written text

• Product recommendations
7 Retail • Product identification and personalisation
• Demand prediction

• Diagnostics
8 Health care • Augmented decision making
• Administration management

• Industrial automation
9 Industry and manufacturing • Intelligent supply chain optimisation
• Product design

• Autonomous vehicles and driver assistance

10 Transport and logistics • Monitoring engine performance
• Enhanced safety and efficacy

• Algorithmic trading
11 Business and finance • Robo-advisors
• Deep data analytics

MACHINE LEARNING INNOVATION - DECEMBER 2019 27

 Figure 8 shows the number of patent families in each category of real world application by their grant
status: granted (where at least one family member is granted in any jurisdiction) or patent application
(a patent family with members seeking patent protection, or in which all members are lapsed/withdrawn
without grant).

These areas showcase current industries where machine learning is being developed and is having an
impact. These industries are – or will soon be – transforming through the use of AI and machine learning
in all aspects of their systems and production. This gives a preview into how the global future of work will
soon change.

FIGURE 8 PATENT FAMILIES RELATING TO MACHINE LEARNING REAL WORLD APPLICATIONS BY THEIR GRANT STATUS
Source: PATSTAT 2018 Autumn Edition

TELECOMMUNICATION 1,997 4,033

ENVIRONMENT AND MINING 1,061 2,434

ENERGY 1,118 2,363

HEALTH CARE 850 2,210

PREDICTIVE MAINTENANCE 710 1,652

INDUSTRY 472 1,077

TRANSPORT AND LOGISTICS 378 1,088

SECURITY AND MILITARY 277 1,062

AGRICULTURE 331 752

BUSINESS AND FINANCE 160 834 Granted patents

RETAIL 189 720 Patent applications

Number of patent families

Telecommunication
The telecommunication sector from home internet network connecting to two or more mobile
is the leader in real world operation, such as streaming communication networks. As the
applications of machine learning. rates and signal strength interference is non-linear, it is
With over 6030 patent families, (WO2017220428 from Alcatel mathematically very expensive to
this sector is maturing fast and Lucent, acquired by Nokia in calculate all possible cases of the
has nearly 2000 granted patents. 201657); and media-specific interference, but the expected
This category includes patent measurement using metrics value of the interference can be
applications for managing mobile such as length and frequency learned and accounted for by
and computer networks, internet of video buffering events and using a multi-layer perceptron.
and data security, as well as audio quality (WO2017144506
This patent family includes a
inventions that allow service from Ericsson).
member granted in the
providers to measure or predict
WO2016040590, filed by United States.
user satisfaction with web-based
Qualcomm, describes a method
services and content. Massachusetts Institute
of removing noise caused
of Technology (MIT) is a
Examples of innovation in this by interference between two
private research university in
sector are the use of metrics channels when a device is

57
Nokia finalizes its acquisition of Alcatel-Lucent, ready to seize global connectivity opportunities, Nokia media release (2 November
2016) https://www.nokia.com/about-us/news/releases/2016/11/02/nokia-finalizes-its-acquisition-of-alcatel-lucent-ready-to-seize-

28
global-connectivity-opportunities/
Cambridge, Massachusetts sensors to measure pollutants Grid controls the world’s largest
and hosts an active group of in air and soil, and to monitor electricity supply network,
researchers working on different and/or predict natural disasters. providing power transmission,
aspects of machine learning power system maintenance and
WO2018072815 (Schlumberger
including theoretical foundations, power network maintenance,
Holdings) is an example of
optimisation algorithms, and as well as other services across
a real world application of
a variety of applications such China. State Grid also has
machine learning in the mining
as vision, speech, healthcare, substantial global holdings in
sector. This invention relates
materials science, NLP and the energy sector, including the
to a method of identifying the
biology.58 WO2016054384 National Grid Consortium of the
location of a subsurface seismic
describes a method to assess Philippines and investments in
layer for optimal drill placement.
the risk of downloading and/ Australia, Brazil, Italy, Portugal
A subsurface apparatus
or using mobile applications, and Greece.
generates a three dimensional
in which operational features
image of the subsurface area Of these 590 patent families,
of mobile applications and
and provides it to a machine five (or less than one per
malware are extracted and
learning process that then cent) have family members
used to train classifiers. The
predicts the presence of a filed in countries other than
assigned classifications and the
subsurface seismic layer in the China. These five international
correlations between them are
three dimensional image. applications by State Grid
used to generate a risk rating.
cover a range of applications,
User-defined security guidelines In environmental science,
and include WO2016141739
can also be incorporated into the US2014099020 (Keimyung
for establishing a supply and
risk classification scheme. University Industry Academic
demand early warning model
Cooperation Foundation)
in the electricity market;
Environment and shows a method to predict
WO2015081660 for forecasting
mining the occurrence of a forest fire
and eliminate false alarms
residential short-term load
using neural networks, and
Environment and mining is a using spatiotemporal Bag of
WO2016033883 for providing
broad category. In the mining Features (BoF) and random
power transmission line gallop
sector, machine learning forest algorithms.
risk early warning prediction
is being applied to mineral using adaptive boosting
exploration, self-driving Energy (AdaBoost) with historical and
vehicles and automation of meteorological data. State Grid’s
mineral processing. This sector Real world applications of
WO2018045642, for predicting
includes 3495 applications, with machine learning in the energy
electrical busbar loading using
30 per cent granted. industry include a range of
random forest classification and
technologies both for generating
Environment and mining real data mining to provide more
energy and for the automated
world applications include accurate decision support for
management of grids. These
inventions that assist in power grid planning and real
technologies maximise output
extracting natural resources, time scheduling, was also filed
from generators, and minimise
and in protecting and monitoring in Australia (AU2016325186), but
the impact of blackouts. The
environments. In the mining lapsed due to lack of acceptance
category is the third largest area
industry, machine learning in examination in early 2019.
of machine learning, with 3481
is being used to assist with patent families, and over 1000 The renewable energy sector
prospecting for minerals, oil patent families granted. is also well represented.
and gas; and automation of Algorithms have been developed
drilling and mineral processing. State Grid have filed 17 per
that use multiple datasets
Applications for environmental cent of patent filings (590
to predict weather patterns
science focus on the use of patent families) in the energy
and forecast the availability
image analysis and remote applications category. State

29
58
MIT Machine Learning Group (accessed 9 October 2019) http://machinelearning.mit.edu/

MACHINE LEARNING INNOVATION - DECEMBER 2019

 of renewable resources,
for solar (WO2016210102,
Predictive Industry
Qatar Foundation) and wind maintenance The industry and manufacturing
(AU2013207551, Tata Consultancy real world applications include
The use of machine learning
Services Ltd; WO2014063436, patent applications for industrial
in predictive maintenance has
State Grid) power. automation, which.
applications across the whole
Energy applications of machine economy. This sector covers The industry category includes
learning also include smaller technologies that are critically technologies of key importance for
scale domestic uses. For important to manufacturing and manufacturing and other industrial
example, WO2018105784, filed other industrial sectors, in which sectors. Patent applications
by the South Korean company machine learning can be used to cover addition of sensors to
Ontest Corporation, describes monitor equipment, structures industrial equipment for optimising
a home energy management or vehicles, and to predict when performance and monitoring,
system using sensors that maintenance or repair work is optimising the supply chain through
record wall temperatures required. These methods can logistical planning, and use of
and use machine learning include the capture and analysis machine learning to help design
techniques to control heating of acoustic signals to predict better products. This category
and cooling of the indoor spaces. when a machine is likely to fail. includes 1549 patent families.
This category contains 2362
Health care patent families, about one third For example, machine learning
programs are being used to
of which have been granted.
The health care category improve workflows in the oil
captures a variety of inventions Clearly, patents relating and gas industry, allowing
that help to deliver better to predictive maintenance users to more easily interrogate
diagnostics and treatment applications can also be applied to information repositories and
for patients and streamline other applications. In this analysis, recommending courses of action
administration of hospitals. we set out to identify applications to improve workflows (for example
There are 3060 patent families that primarily related to predictive WO2016153470 from Haliburton
for real world machine learning maintenance systems as a priority. Energy Services Group).
applications in healthcare. An example patent application An example in this category is a
Technologies include using related to predictive maintenance patent application that describes
machine learning to assist with is WO2015070513 (State Grid). This the use of sensors and data
medical imaging (e.g. reading patent application describes the analysis to measure the health
magnetic resonance imaging use of machine learning employed of structures. National ICT
(MRI) scans) and biological to detect patterns in insulation Australia (NICTA) (now part of
sample testing. defects related to gas insulated Commonwealth Scientific and
Machine learning is also switchgear (GIS) used in ultra-high Industrial Research Organisation
augmenting medical decision voltage grids. Pattern recognition (CSIRO)) (WO2014176625)
making by predicting how allows control of partial discharge measures vibrations caused
a patient’s condition will of power, which prevents electrical by traffic for structural health
change based on their medical breakdown or failure. monitoring for bridges. Ambient
history and what side-effects factors such as wind are
More broadly, WO2017213810
a particular treatment could measured, and machine learning
(Exxonmobil Research and
cause (WO2017019706, held by is used to determine the health
Engineering Company) describes
Alphabet, the parent company of the structure. This cross
the use of machine learning
of patent applicant Google), cutting technology is also relevant
techniques to detect and analyse
and potential risks associated to other applications such as
visual and audio events.
with genetic variants identified predictive maintenance and
This information can be used to
through genomic screening transport and logistics.
control a range of automated
(WO2016209999, Counsyl). industrial processes.

30
CASE STUDY: MACHINE LEARNING-ASSISTED MEDICAL IMAGING
A large amount of diagnostic specialised pattern recognition that may cause abnormalities
effort across most fields tasks, vast data streams and in the images in consideration
of medicine relies on the a corpus of tagged training (for example osteoporosis in
interpretation of various forms data means that medical bone images). This combination
of imaging. Image analysis imaging is an area well suited of patient-specific information
is key to areas of medicine to enhancement by machine with medical image information
from cancer detection and learning. As well as this, is expected to give more
characterisation (both internal improvements in medical reliable information than image
and on the skin) to fracture imaging are very important to information alone.
evaluation, diagnosis of both patients and to society. Australia has a technology
infection, circulatory studies, Enhancement of the capabilities strength in medical technologies
blood clot and aneurism of a limited supply of medical in general, and medical
detection and many more. professionals promises to be imaging in particular. An
Machine learning is already one of the most valuable and example of an Australian patent
applied to medical imaging rapidly adopted applications of application comes from the
problems, including machine learning. CSIRO (WO2016019439). This
ophthalmology and skin lesion Patenting in this area is active, patent application relates to
detection,59 and research with more than 1250 families imaging the interior of a body
suggests this market will grow filed since 2012 and 403 in 2017 cavity, with a specific use in
to $2 billion USD by 2023.60 alone. The global leader in effectively detecting colorectal
Anticipation of this emerging medical imaging patenting is abnormalities. A machine
market has led large computer the German company Siemens learning system is trained with
manufacturers such as graphics AG with 86 patent families. Their real image data and produces
processing specialist NVIDIA to patent US2014185888 relates simulated images that can
release purpose built platforms to non-invasive lesion detection be used to generate models.
for medical image applications.61 from medical images such as These models can be used for
The medical profession stores computed tomography (CT) and a variety of purposes including
data that is highly detailed, MRI. Their technique consists of for training purposes (reducing
systematically archived and two major components. Image- the burden on real patients) and
classified by professionals related information such as for enhancing the capability of
with years of professional absorption and geometry are clinicians by showing in real
experience and training. This collected and interpreted. The time any areas that have not
gives rise to a large amount analysed images are examined been examined.
of pre-tagged ground-truth together with patient medical
data. The combination of highly history data regarding conditions

59
How AI and Deep Learning Are Revolutionizing Medical Imaging, GEHealthCare (08 April 2019)
https://www.gehealthcare.com/long-article/how-ai-and-deep-learning-are-revolutionizing-medical-imaging
60
Harris, S (16 October 2018) What’s new for machine learning in medical imaging, Signify Research,
https://www.signifyresearch.net/medical-imaging/signify-ai-medical-imaging-white-paper/
61
Salian, I (4 March 2019) How AI is changing medical imaging, NVIDIA blogs,

31
https://blogs.nvidia.com/blog/2019/03/04/how-ai-is-changing-medical-imaging/

MACHINE LEARNING INNOVATION - DECEMBER 2019

 Transport and received by a smart device. It
identifies the characteristics of
by Facebook for a face tracking
system that predicts face
logistics a person’s lip motion to detect alignment by using facial anchor
individual words and uses deep points to create a mesh shape
The transport and logistics area
learning methods alongside and to track facial movement
is active in the development of
sound wave signal processing (Figure 9). This system reduces
driverless vehicles, from cars
technology to model and predict to a model that can be used
to trains and aeroplanes, and
sounds CN107784215. on devices to predict facial
sensors for monitoring engine
movement WO2018118099.
performance and predicting the An example of facial scanning
need for maintenance. It also includes a patent application
includes artificial intelligence
used in space technologies and
spacecraft. This category has FIGURE 9 IMAGE FROM FACEBOOK’S PATENT APPLICATION WO2018118099
1466 patent families.

A patent application filed by

Netposa Technologies Ltd is an
example of an image analysis
systems that uses deep learning
techniques to read license plate
numbers (WO2018028230).

Improving safety is an area

of commercial interest. For Cropped Barycentric Mesh-
example, Uber Technologies Bounding Box based Default Shape
have filed patent application 166 210
WO2018073662 for a travel
safety system that collects data,
such as dangerous driving or
interpersonal conflict incidents,
to predict future safety incidents.

Security and
military
The security and military
category covers a broad range
Adjusted Barycentric Determined Facial
of inventions, including facial
Mesh-based Default Anchor Points by
recognition, biometric scanning Default Shape
Shape
(facial, fingerprint and voice 220 230
recognition), fraud detection,
reading sentiment in written
text, and defence applications.
This category includes 1339
patent families.

Examples of the use of these

technologies include a patent
application by Shanghai Jiao
Barycentric Mesh-
Tong University for a system
based Fitted Shape Determined Facial
to lip reading based on sound 240 Anchor Points by 240
250

32
MACHINE LEARNING INNOVATION - DECEMBER 2019 33
 Agriculture developed by Intel (US9898688).
This system uses neural
credit card purchase special
offers that are aligned with
In agriculture, machine learning networks to analyse and classify those interests (WO2014204693).
is being used in robots that agricultural conditions based Personalised recommendations
reduce the cost of field labour, on data collected by drones and are also important in emerging
using image analysis to monitor generates a three dimensional inventions in entertainment,
crops and soil quality, and for map showing agricultural as are virtual and augmented
quarantine and food inspection. conditions. This method is reality for games.
sensitive enough to distinguish
Agriculture machine learning In finance, machine learning
between crop disease, crop
real world applications reduce is being used in business
damage caused by animals and
the cost of farm labour through investments, such as algorithms
water damage.
AI-controlled field robots, use for making trades, automated
assistants for managing
of imaging to monitor soil and
crops, use of predictive analytics
Business and portfolios and assessing risk,
tools to maximise production, finance and predictive analytics.
and also development of
The business and finance Patent application
improved quarantine and food
category covers patent WO2018074902 by Samsung
inspection processes. We found
applications relating to AI used describes a system for mobile
1083 patent families relating to
in business investments, such payments that, when a request
the use of machine learning in
as algorithms for making trades, to send money from a sender to
agriculture, indicating that this
AI assistants for managing a receiver is made, a computer
area has strong potential growth.
portfolios and assessing risk, program generates a list of
An example of the use of fraud detection, and predictive eligible providers based on the
machine learning in water analytics. This category has transaction, and then notifies a
resource use is from 994 patent families. The provider to transfer the money.
CN107578116 (National Dong business and finance category
Fraud detection is vital to a viable
Hwa University, Taiwan). This brings together a range of
business and finance sector.
invention aims to improve the technologies involved in finance,
Patent application US2017200164
distribution of agricultural water entertainment, and education.
(Korea Internet and Security
resources during drought using
Business applications include, Agency) is an example of
an immune particle swarm
for example, the analysis of machine learning in detection of
optimisation algorithm. This
customer behaviour. Machine fraudulent transactions.
maximises crop yields even in
learning is used to analyse
drought conditions.
customers’ credit card
Vision enhanced drones for transactions to determine their
precision farming have been interests, and then recommends

34
CASE STUDY: PROPERTY VALUATION
A niche application of machine and Redfin provide AI-assisted training data from the same
learning techniques is in the home valuations as part of their geographical area.
valuation of property. Property database offering.66 Recently the Patent CN107230113 from the
is big business in Australia. University of New South Wales has Jiangxi University of Science and
According to some sources the developed a suite of automated Technology is also related to the
total value of the Australian land valuation software tools valuation of a property based
residential property market alone called Value Australia.66 on information for recorded
is more than $7.5 trillion AUD.62 While patenting activity appears transactions for similar properties.
Property valuation is difficult relatively muted in property Ongoing activity in property
due to high variability between valuation overall, some patent valuation seems poised to enable
properties. Properties must applications from across the world rapid, accurate and often free
also be valued at different times were identified in this analysis. valuations of property, both
and by different parties, not only Examples included US2018068329 residential and commercial,
at the time of sale but also for from IBM relates to automated resulting in useful outcomes
assessments such as land taxes. property valuation using neural for property investors as well as
There are many online AI tools networks. This uses images at those who need information about
available for valuing residential various levels of zoom to see the property market, such as
and commercial properties.63,64 features of the property and its government agencies.
In the United States, real estate surrounding area to predict the
databases such as Zillow, Trulia price of the property based on

Retail learning to generate product

recommendations, chatbots for
techniques to analyse information
about a user to provide targeted
The real world applications of customer assistance, and also for advertisement selection and
machine learning in retail, with 909 the prediction of product demand display, which assists customers
patent families identified, includes and supply management. as well as improving sales.
inventions where machine learning
The biggest retail players are Overall, this section on real
is used to create personalised
well represented here. Ebay world applications is exciting and
online product recommendations,
holds patent applications demonstrates the diversity and
personalised advertising and
such as WO2018089762, for significance of developments
customer engagement. It
personal assistance with online in machine learning across all
covers all web-based shopping
shopping. This system leverages areas of our lives and economy.
and e-commerce platforms. It
existing inventories and curated In the next section, we will take
also captures applications for
databases to provide intelligent a closer look at the development
predicting product demand.
answers to a human user, using of foundational machine learning
The retail category captures image analysis including text-in- technologies that are allowing
technologies that are being image searching. these changes.
used to increase customer
Yahoo (US2016125314) describes
engagement and improve
the use of machine learning
sales. This includes machine

62
Yardney, M (21 December 2018) State by state: An end-of-year update on Australia’s property markets, Smart Company,
https://www.smartcompany.com.au/industries/property/end-year-update-australia-property-markets/
63
Chagani, E (28 March 2018) The potential of machine learning real estate valuation models, Cornell University,
https://blog.realestate.cornell.edu/2018/03/28/machine-learning/
64
Wiggers, K (24 January 2019) Geophy raises $33 million to apply AI to commercial property appraisal, Venture Beat,
https://venturebeat.com/2019/01/24/geophy-raises-33-million-to-apply-ai-to-commercial-property-appraisal/
65
Tarver, E (17 April 2019) The 5 Best Alternatives to Zillow and Trulia, Investopedia,
https://www.investopedia.com/articles/markets/100215/5-best-alternatives-zillow-trulia.asp
66
Johnston, M (4 October 2019) UNSW partners with govt, industry for AI land valuation, ITNews

35
https://www.itnews.com.au/news/unsw-partners-with-govt-industry-for-ai-land-valuation-531896

MACHINE LEARNING INNOVATION - DECEMBER 2019

06
Foundation
technologies
Now we have considered the applications of machine learning – both core
capabilities and real world applications, we will turn to the foundational
technologies that support the continued development of machine learning.
We identified 24,091 patent Overall 26 per cent of the patent The architecture category
families relating to machine families relating to the machine includes patent families that
learning foundational learning techniques have at least describe physical realisation
technologies, and divided these one family member granted. or hardware implementation
into two categories: algorithms of neural networks, neural
The algorithms category includes
and architecture (see Appendix network architecture and its
patent families that describe
C: Technology Analysis for more optimisation. Only 15 per cent of
supervised, semi-supervised
information about the sub- the patents families in machine
or unsupervised learning,
technology classification). learning techniques relate to
reinforcement learning, other
this category, with 3,600 patent
Figure 10 shows the number niche learning methods such
families, of which 20 per cent
of patent families in each as ensemble learning and
have a granted family member.
category by their grant status: extreme learning, and general
granted (where at least one machine learning techniques. In the next section of this report,
family member is granted in any This category accounts for 85 per we will take a closer look at
jurisdiction) or patent application cent of patent families relating each of these categories and
(a patent family with members to machine learning techniques, their top applicants.
seeking patent protection, or in totalling 20,491 patent families,
which all members are lapsed/ of which 30 per cent have a
withdrawn without grant). granted family member.

FIGURE 10 PATENT FAMILIES RELATING TO MACHINE LEARNING FOUNDATIONAL TECHNOLOGIES BY THEIR

GRANT STATUS
Source: PATSTAT 2018 Autumn Edition

ALGORITHMS 5,527 14,964

ARCHITECTURE 706 2,894

Granted patents Patent applications

36
Top applicants
The number of patent families filed by an applicant in a specific
technology can be indicative of their interest and market presence
or desire to build and maintain a market share. Figure 11 displays
the top applicants by their patent families on machine learning
foundational technologies.

FIGURE 11 PATENT FAMILIES RELATING TO MACHINE LEARNING FOUNDATIONAL TECHNOLOGIES BY APPLICANT

Source: PATSTAT 2018 Autumn Edition

Algorithms

STATE GRID 720 11

CHINESE ACADEMY OF 611 39

SCIENCES

XIDIAN UNIVERSITY 426 7

ZHEJIANG UNIVERSITY 331 19

UESTC 300 19

Architecture

IBM 71 47

QUALCOMM 40 59

ALPHABET 74 21

CHINESE ACADEMY OF 58 18
SCIENCES
Foundation technologies and applications
SAMSUNG 48 24 Foundation technologies only

MACHINE LEARNING INNOVATION - DECEMBER 2019 37

 The top five filers of patent AI-related smart algorithms Their portfolio also includes
applications relating to and robotics. State Grid’s grid inventions relating to neural
algorithms are Chinese. operation and management network architecture and
Most patent families filed by systems involve the collection implementation. In 2017, Xidian
these applicants describe and analysis of various University established the School
both algorithms and their data, using AI technologies of Artificial Intelligence, which
applications, which may be an including image processing, focuses on innovative research
indication of interest in applying voice recognition and big data and development in AI, including
machine learning techniques to analysis.69 State Grid is one of fields such as intelligent
real world applications. the two state-owned enterprises information processing,
that operate the power network intelligence and image
Patent filings relating to
in China, the other being China understanding, visual computing
architecture are led by American
Southern Power Grid (CSPG), and collaborative cognition.74
applicants. Architecture
which has a service area
patent applications show a Zhejiang University is a public
spanning one million square
greater proportion of filings in university located in Zhejiang,
kilometres and a population of
techniques development only. China. Their patent portfolio
252 million people.70 CSPG is also
predominantly includes
State Grid is the top applicant an active player in the machine
algorithms related inventions and
relating to algorithms, with a learning field with a total of 263
they rank fourth worldwide in this
total of 763 patent applications in patent families filed since 2012.
category. The university’s Institute
either foundational technologies
The Chinese Academy of of Artificial Intelligence focuses
or foundational technologies and
Sciences is active in both on research areas including AI
applications. State Grid is the
algorithms and architecture, theory, computer aided design
world’s second largest company
and hosts 104 research and computer graphics, client
on the Fortune Global 500 list,67
institutes, including the Institute information management
and is the largest public utility
of Intelligent Machines.71 This systems, and other advanced
company in the world, supplying
institute has become one of manufacturing technologies.75
power to over 1.1 billion people in
the leading national research
88 per cent of Chinese UESTC ranks in fifth place for
centres in sensor technology
national territory.68 filing algorithms related patent
and AI. The research centre
applications. This is a public
State Grid focuses on AI focuses on biomimetic sensing
university located in Sichuan
developments in power grid and intelligence related to
province, China. In 2014, UESTC
control, power distribution and robotics and human-computer
established an AI research
utilisation networks, and green interactions, aiming to advance
institute called the Statistical
energy. Their patent portfolio digital agriculture, sensor
Machine Intelligence and
is focused on application of performance, control systems
Learning (SMILE) Lab. The lab
machine learning techniques, and administration.72
studies algorithms for solving
especially in energy and
Xidian University is a public various data analytics problems
predictive maintenance. The
research university located in relating to AI and machine
company also contributes to
Xi’an, China,73 ranking third in learning.76 Their patent portfolio
the development and use of
algorithms related inventions. also includes inventions relating

67
Global 500 (accessed 9 October 2019) http://fortune.com/global500/
68
State Grid Corporation of China, Corporate Profile (accessed 9 October 2019)
http://www.sgcc.com.cn/html/sgcc_main_en/col2017112307/column_2017112307_1.shtml
69
China Institute for Science and Technology Policy at Tsinghua University (July 2018) China AI Development Report 2018,
http://www.sppm.tsinghua.edu.cn/eWebEditor/UploadFile/China_AI_development_report_2018.pdf
70
China Southern Power Gridgrid, Company Profile (accessed 9 October 2019)
http://eng.csg.cn/About_us/About_CSG/201601/t20160123_132060.html
71
Chinese Academy of Sciences, CAS Institutes (accessed 9 October 2019) http://english.cas.cn/institutes/
72
Chinese Academy of Sciences, Introduction for the Institute of Intelligent Machines (accessed 9 October 2019)
http://english.hf.cas.cn/r/ResearchDivisions/IIM/
73
Xidian University, Overview (accessed 9 October 2019) https://en.xidian.edu.cn/About_Xidian/Overview.htm
74
Xidian University, School of Artificial Intelligence (accessed 9 October 2019) https://en.xidian.edu.cn/info/1018/1945.htm

38
75
Zhejiang University, Institute of Artificial Intelligence (accessed 9 October 2019)
http://fit.zju.edu.cn/english/redir.php?catalog_id=2443
76
University of Electronic Science and Technology of China (accessed 9 October 2019) https://en.uestc.edu.cn/
to neural network architecture architecture and implementation. categories, and ranks third for
and implementation. More of their inventions inventions relating to neural
focus on the developments in network architecture
International Business Machines
machine learning foundational and implementation.
Corporation (IBM), the top
technologies than on applications
applicant in neural network Samsung, a South Korean
of machine learning.
architecture and implementation, multinational conglomerate,
is an American computer and Alphabet is an American is developing technologies
information technology consulting multinational conglomerate including AI, data intelligence,
company. The company’s that owns multiple subsidiary intelligent machines and next
research centre, IBM Research, companies including Google, generation communication, with
explores AI and machine learning Waymo and DeepMind the aim of creating cutting-edge
techniques, with a focus on Technologies.79 Google already AI technologies for application in
three areas: advancing AI, uses machine learning to power all of its products and services.84
scaling AI, and trusting AI.77 The their products including search Samsung ranks in fifth place for
company’s patent profile includes engines, Google maps, YouTube, patent applications relating to
fundamental developments in Gmail and Google drive, with neural network architecture and
machine learning techniques new techniques leading the way implementation. They also have
as well as applications of from mobile-first to AI-first in the filed patent applications relating
machine learning techniques in world.80 Waymo, Google’s self- to algorithms and real world
telecommunication and networks, driving car, is committed to the application of machine
retail, medicine and healthcare. use of applied AI and machine learning techniques.
learning techniques to make
Qualcomm is an American
telecommunication technology
roads safer and improve mobility Algorithms
for everyone.81 DeepMind focuses
company, with a focus on Foundational technologies
on developing algorithms that
creating, developing and related to algorithms are
can learn to solve any complex
commercialising power- sub-divided hierarchically into
problem without needing to be
efficient on-device AI to make supervised and semi-supervised
taught how.82 One success story
mobile phones a pervasive AI learning, unsupervised learning,
is DeepMind’s AlphaGo, the first
platform.78 In this analysis, reinforcement learning, niche
computer program to defeat a Go
Qualcomm is the second learning techniques and general
world champion.83 Alphabet has
ranked applicant for inventions machine learning techniques
a strong patent portfolio across
relating to neural network (see Figure 12).
the algorithms and architecture

FIGURE 12 PATENT FAMILIES RELATING TO ALGORITHMS

Source: PATSTAT 2018 Autumn Edition

SUPERVISED AND SEMI-SUPERVISED LEARNING 14,086

UNSUPERVISED LEARNING 3,283

NICHE LEARNING TECHNIQUES 1,999
GENERAL TECHNIQUES 648
REINFORCEMENT LEARNING 475

77
IBM Research (accessed 9 October 2019) https://www.research.ibm.com/artificial-intelligence/
78
Qualcomm Research & Invention (accessed 9 October 2019) https://www.qualcomm.com/invention/artificial-intelligence
79
Alphabet, G is for Google (accessed 9 October 2019) https://abc.xyz/
80
AI Multiple (21 february 2019) Google is AI first: 12 AI projects powering Google products,
https://blog.aimultiple.com/ai-is-already-at-the-heart-of-google/
81
AI Multiple (11 December 2018) Alphabet’s AI Investments in 38 Companies: In-depth Analysis, https://blog.aimultiple.com/alphabet-ai/
82
DeepMind, About Us (accessed 9 October 2019) https://deepmind.com/about/
83

84
AlphaGo (accessed 9 October 2019) https://deepmind.com/research/alphago/
Samsung Research Areas (accessed 9 October 2019) https://research.samsung.com/research-areas 39
 Supervised and semi- fall into other categories of areas including computer vision
supervised learning is the machine learning foundational systems, target tracking, data/
largest sub-category, with a technologies or machine learning document clustering, credit risk
four-fold increase in patent applications (see Appendix C). analysis, speech recognition and
family filings between 2012– bioinformatics. Patent protection
The following patent
16. This category includes for this patent family is being
applications are examples
classification and regression sought in the United States.
of algorithms applications in
techniques, supervised
these sub-categories. Patent application
dictionary learning, supervised
US20180129974 discloses a deep
neural networks (including Patent application CN102831404,
reinforcement learning technique
supervised deep learning) and filed by the Chinese Academy of
relating to control systems. The
boosting algorithms. Sciences, describes a gesture
technique determines an action
detection method using machine
The second largest sub- (for example the control of speed
learning classifiers (supervised
category is unsupervised and angle of a nozzle in a material
learning) on both colour and
learning, including cluster deposition process) based on
grayscale versions of an image.
analysis techniques, blind environment observations (e.g.
The method employs a sliding
signal separation techniques, LiDAR or video images). The
window to extract features from a
unsupervised dictionary learning, outcome of the action is then
region of an image, an adaboost
and unsupervised neural assessed against the desired
classifier for classifying features,
networks (including unsupervised outcome and this is fed back
and a random forest classifier for
deep learning, autoencoders, self- into the control system to
classifying point-pair features (a
organising maps and GANs). inform the choice of subsequent
type of image feature that is used
actions, along with continued
Reinforcement learning accounts for rigid object detection in point
environmental observation.
for two per cent of the patent cloud data). This patent family has
The deep reinforcement
families in algorithms and has a member granted in China.
learning technique uses a
also increased four-fold in five
An patent application describing convolutional neural network
years. This sub-category includes
an unsupervised learning or a deep autoencoder. This
dynamic programming, learning
technique is WO2015088780, patent application was filed by
automata and reinforcement
by the University of Southern United Technologies Corporation,
neural networks.
California, which hosts the an American aerospace and
The niche learning techniques Center for Artificial Intelligence building technology company with
sub-category includes patent in Society for conducting headquarters in Connecticut.86
families that discuss ensemble research in AI to help solve Their research centre focuses
learning, extreme learning, difficult social problems facing on developing machine learning
transfer learning, genetic the world.85 This application techniques for intelligent systems
algorithms (with machine proposes a noise enhanced that allows for self-diagnosis and
learning), rule learning, swarm clustering technique, consisting repair.87 Patent protection for this
optimisation, online learning, of applying an iterative clustering invention is being sought in the
feature learning, graphical method (k-means clustering United States and Europe.
models, mixture models and algorithm or competitive learning
relational learning. algorithm) along with artificial
noise, which is used to speed up
The general techniques sub-
convergence and increase the
category contains patent families
robustness of the final clusters.
that mention machine learning
This clustering technique may
or neural networks and do not
find application in a range of

85
USC Center for Artificial Intelligence in Society (accessed 9 October 2019) https://www.cais.usc.edu/
86
United Technologies (accessed 9 October 2019) https://www.utc.com/

40
87
United Technologies Research Center (accessed 9 October 2019) http://www.utrc.utc.com/platform-technologies.html
CASE STUDY: TRANSFER LEARNING – THE NEXT EDGE OF MACHINE LEARNING
If AI is the simulation of human using unsupervised transfer this method helps reduce
intelligence processes by learning, and reinforcement the number of experiments
machines, then machine learning techniques using transductive required to determine optimal
could be considered a simulation transfer learning) and is control parameters, by
of human learning by machines. divided across the categories leveraging knowledge gained
Humans don’t learn everything in Figure 12, including niche from past experiments to
from scratch. We leverage our learning techniques where adjust control parameters in
previous experiences and apply supervision is not specified. subsequent experiments.
earlier information to new tasks Methods of transfer learning General Electric Company90
and environments. Inspired by can be interchangeably used has filed patent application
this process, transfer learning in these categories between US2017300605, describing the
leverages and transfers models in different domains. use of transfer learning for
knowledge acquired from These methods include instance predictive maintenance. This
solving one task in one domain transfer, feature representation application describes a method
to solve other tasks in other transfer, parameter transfer and for creating a predictive damage
domains.88 Transfer learning relational knowledge transfer. model for aviation assets, by
marks a paradigm shift in Our analysis shows dramatic predicting behaviours of a
machine learning, from models growth in filing trends for given physical asset and then
designed to learn in isolation to transfer learning from eight using transductive transfer
models designed to learn from patent families in 2012 to 53 learning from this model (with
shared past knowledge. patent families in 2017, with labelled data only available
For example, a model trained the number of filings doubling in the first model) to create
using traditional techniques between 2016–17. Transfer a second predictive damage
to detect pedestrians in night- learning is widely used with model to predict the behaviour
time images would fail to detect unstructured data and this is of a second physical asset in a
pedestrians on a day-time reflected in its application. Of separate domain, such as new
image and would require a new the 134 transfer learning patent or expanded asset types or new
model to be trained separately families, 54 per cent are used flight conditions.
with day-time images. Transfer for image and video analysis Another example is granted
learning can instead leverage while 19 per cent relate to patent US10007863 by applicant
the knowledge gained from speech and text analysis. Gracenote, a subsidiary of Nielsen
training on night-time images Examples of patent applications Holdings. This patent describes a
and apply it to training for on transfer learning include method for accurately detecting
detecting pedestrians in day- WO2017173489, originating logos in media content on media
time images, thereby using from Australia and filed by presentation devices using neural
fewer training images and Deakin University, which networks and transfer learning.
reducing the time to train. hosts an AI and Data Analytics As predicted by the Chief
Transfer learning can be applied Research Cluster.89 This Scientist at Baidu and
to machine learning techniques patent application describes professor at Stanford
using algorithms designed for a method for calculating University, Andrew Ng, transfer
each type of technique (i.e. to control parameters to make learning is becoming the next
supervised techniques using an optimised manufactured driver of machine learning
inductive transfer learning, product. Designed for use in commercial success after
unsupervised techniques early product development, supervised learning.91

88
Pan, S. J. and Yang, Q., ‘A Survey on Transfer learning’, IEEE Transactions on Knowledge and Data Engineering, 22 (10) Oct 2010
https://ieeexplore.ieee.org/document/5288526
89
Deakin University Artificial Intelligence and Data Analytics Research Cluster (accessed 9 October 2019)
https://www.deakin.edu.au/information-technology/research/artificial-intelligence-and-data-analytics
90
Company Information (accessed 9 October 2019) https://www.ge.com/news/company-information
91
Kim B-H (6 May 2018) NIPS 2016 tutorial: ‘Nuts and bolts of building AI applications using Deep Learning’ by Andrew Ng,

41
https://www.youtube.com/watch?v=wjqaz6m42wU

MACHINE LEARNING INNOVATION - DECEMBER 2019

 Architecture
Foundational technologies related to architecture are sub-divided into
neural network architecture, physical realisations, and architecture
optimisations (Figure 13).

FIGURE 13 PATENT FAMILIES RELATING TO ARCHITECTURE

Source: PATSTAT 2018 Autumn Edition

2,536
NEURAL NETWORK ARCHITECTURE

PHYSICAL REALISATIONS 982

ARCHITECTURE OPTIMISATION 82

Neural network architecture is that maps a neural network Patent application

the largest sub-category, with onto a neurosynaptic substrate WO2016123409, originally filed
70 per cent of patent families. (a computing architecture by Google, relates to a neural
This sub-category includes specifically optimised for use with network system including
patent families describing the neural network algorithms) based a normalisation step (a
architectural layout of neural on information about a graph (or ‘normalisation layer’) between
networks and how the neurons ‘adjacency matrix’) representation successive layers of the neural
are interconnected. of that neural network. Users can network. When the neural
interact with the mapping system network is being trained, the
Physical realisations are the
in order to trade-off between normalisation layer receives
second largest sub-category,
accuracy and resource utilisation the output of a given layer of
with 27 per cent of patent families
of the substrate. This patent the neural network for each of
in the architecture category.
family has a member granted in the training examples in a given
This sub-category includes
the United States. batch, and then normalises
implementation techniques
these outputs based on the
used to physically construct the Qualcomm has filed
statistics of the batch as a
neurons or neural networks. WO2015038311, a patent
whole, before providing the
application describing
Architecture optimisation normalised output to the next
optimisation of a neural network
includes patent applications that layer of the neural network.
system in which computational
describe optimisation of neural The application claims this
resources are limited by
network architecture by adding technique increases the training
structural plasticity. Structural
or deleting nodes or connections. speed of a neural network, and
plasticity is the ability for
This sub-category represents only can also reduce the need for
elements of a neural network to
two per cent of patent families in regularisation (mechanisms to
change their shapes and thereby
the architecture category. restrict the degree of overfitting
connections in the network. This
of a given model). This patent
The following patent invention alters the location of
family has members granted in
applications are examples of synapses connecting particular
Europe, Australia and Spain.
architecture applications in neuron pairs based on timing
these sub-categories. information from previous
executions. One family member
Patent application WO2016146278, of this patent family has been
by IBM, describes a system granted in the United States.

42
MACHINE LEARNING INNOVATION - DECEMBER 2019 43
07
Australian
Applicants
To better understand Australian patenting activity,
we analysed filings by Australian applicants.

This report captures 59 patent further analysis of the filings by place is HRO Holdings with
families filed by Australians, with Australians, by application type. three patent families. In shared
the largest growth between 2015 third place are four entities:
Australian applicants favour
(10 patent families) and 2016 (19 Atlassian, CRC Care, NewSouth
analytics and processing
patent families). Australia is the Innovations and the University
(41 per cent) and image and
primary patent destination for of Technology Sydney, with each
video analysis (31 per cent) in
Australian applications, with 54 having two patent families.
the development of machine
per cent of families filed here,
learning core capabilities. Apart from the top five
followed by the United States
Australian applications, all other
with 34 per cent. Figure 14 shows Patent applications in real world
Australian applicants in our
the divide across applications applications of machine learning
data set have each filed a single
and foundational technologies. by Australian applicants are filed
patent application.
across all technology areas, with
Applications of machine
a focus on health care (Figure The top applicants vary greatly
learning are clearly the focus
16), which is an Australian in size and in their technology
for Australians. Figure 14
technology strength. focus, from multi-billion dollar
shows 97 per cent of patent
companies to universities,
families discuss an application Figure 17 shows the top five
publicly funded research
and only 44 per cent discuss a Australian applicants. The
organisations and start ups.
technique, including a 40 per top patent filer is CSIRO, with
cent overlap. Figure 15 shows a five patent families. In second

44
FIGURE 14 TECHNOLOGY OVERVIEW OF MACHINE LEARNING PATENT FAMILIES BY AUSTRALIAN APPLICANTS
Source: PATSTAT 2018 Autumn Edition

33

24

Applications only

Foundation technologies only

Applications and Foundation technologies

FIGURE 15 PATENT FAMILIES BY AUSTRALIAN APPLICANTS BY CORE CAPABILITIES

Source: PATSTAT 2018 Autumn Edition

ANALYTICS AND PROCESSING 20

IMAGE/VIDEO ANALYSIS 15

SPEECH/TEXT ANALYSIS 6

SIGNAL ANALYSIS 4

CONTROL AND AUTOMATION 3

FIGURE 16 PATENT FAMILIES BY AUSTRALIAN APPLICANTS BY REAL WORLD APPLICATIONS

Source: PATSTAT 2018 Autumn Edition

HEALTH CARE 11

TELECOMMUNICATION 7

ENVIRONMENT AND MINING 6

BUSINESS AND FINANCE 5

INDUSTRY 4

RETAIL 2

ENERGY 2

TRANSPORT AND LOGISTICS 1

SECURITY AND MILITARY 1

PREDICTIVE MAINTENANCE 1

AGRICULTURE 1

MACHINE LEARNING INNOVATION - DECEMBER 2019 45

 CSIRO
CSIRO is Australia’s national science research agency and largest patent
holder, with more than 1800 patent applications across all technologies.92
CSIRO is Australia’s largest public research organisation, with around 5
500 employees and funding of approximately AU $600 million in 2017.93
In November 2018, CSIRO invested AU $19 million in frontier research in
AI and machine learning to target AI-driven solutions for areas including
health and wellbeing, sustainable energy and resources, and Australian
and regional security.94 CSIRO filed machine learning patent applications
related to social media text analysis (WO2017106904, as NICTA) data
visualisation, privacy preserving machine learning WO2016061628,as
NICTA) image analysis (WO2014201521), medical imaging and diagnostics
(WO2016019439) and structural health monitoring (WO2014176625, as
NICTA) from 2013–15.

HRO HOLDINGS
Founded in 2015, inventor Jock Lawrie - through Australian company HRO
Holdings95 - has filed three machine learning patent families in 2016, in
the areas of patient procedure scheduling and outpatient clinic attendance
(WO2018058189, WO2017120646 and AU2017202982).

ATLASSIAN
Founded in 2002, Atlassian is an Australian success story. This
software company offers products for software developers, project
managers and other business users.96 Atlassian is the first ‘unicorn’
start-up company (a start-up valued at more than US $1 billion) to
emerge from Australia.97 The company’s products include Loom,
an AI-powered log analysis solution that informs the user of any
problems within a digital system; Talla, an AI-powered information
technology service management assistant; and Jira, a service desk
software with automation systems.98 They filed machine learning
patent applications related to managing conversations in messaging
applications (US9804752) and the return of user-specific content in
searching (US2018189282) in 2016.

92
CSIRO, Our mission (accessed 9 October 2019) https://www.csiro.au/en/About/We-are-CSIRO
93
CSIRO 2017-2018 Annual report (accessed 9 October 2019)
https://www.csiro.au/~/media/About/AnnualReport/Files/2017-18/18-00410_CORP_AnnualReport2017-18_WEB_180928.
pdf?la=en&hash=38A094CABD16B4D934CD44C309B8F95CB700F78E
94
CSIRO News Release (19 November 2018)
https://www.csiro.au/en/News/News-releases/2018/CSIRO-invests-35M-infuture-of-space-and-AI-for-Australia
95
ASIC Connect search results for HRO Holdings Pty Ltd (accessed 9 October 2019)
https://connectonline.asic.gov.au/RegistrySearch/faces/landing/SearchRegisters.jspx?_adf.ctrl-state=xixewdp8p_4
96
Atlassian, Overview (accessed 9 October 2019) https://www.atlassian.com/software/jira/guides/getting-started/overview
97
Noyes, K (11 December 2015) Atlassian president on IPO - ‘We’re overjoyed’,
https://www.pcworld.idg.com.au/article/590624/atlassian-president-ipo-we-re-overjoyed/

46
98
Atlassian, Software (accessed 9 October 2019) https://www.atlassian.com/software
CRC CARE
Founded in 2005, the Cooperative Research Centre for Contamination
Assessment and Remediation of the Environment (CRC CARE),
performs research, develops technologies and provides policy
guidance for assessing, cleaning, and preventing soil, water and air
contamination.99 Their partner organisations include many of the
world’s largest resource companies such as BHP Billiton, Rio Tinto
and Chevron, as well as universities across Australia, government
bodies and CSIRO.100 In 2015, CRC CARE filed machine learning patent
applications related to determining the concentration of ions in liquids
(WO2016145481) and for determining the presence of volatile organic
compounds in samples (WO2016187671).

NEWSOUTH INNOVATIONS
NewSouth Innovations Pty Ltd is the commercialisation arm of the
University of New South Wales, one of Australia’s largest universities
and one of the Group of Eight research intensive universities,
with laboratories devoted to computer vision, AI and robotics.101,102
NewSouth Innovations filed machine learning patent applications
related to computer security (WO2014094034) and real time moving
image recognition (WO2014194345) in 2012 and 2013, respectively.

UNIVERSITY OF TECHNOLOGY SYDNEY

The University of Technology Sydney established a research centre, the
Centre for Artificial Intelligence (CAI), in 2017. CAI is a world leading
research centre in AI, with a vision to develop theoretical foundations
and advanced algorithms for AI and to drive significant progress in
related areas including machine learning, business intelligence,
computer vision, data science and social robotics.103 The University of
Technology Sydney filed machine learning patent applications related to
traffic lane detection (US9286524) and object tracking (US2016342837),
both co-filed with Toyota in 2015.
Overall, this data shows Australia is a minor player in machine learning
patenting globally. The picture of Australian patent filings in machine
learning is one of diversity, both in the type of applicants, and in the
subject matter of their innovation. This highlights a broad range of
experience here, and indicates strong potential for development of
Australian research in this field.

99
CRC CARE, About (accessed 9 October 2019) https://www.crccare.com/about
100
CRC CARE, Partners (accessed 9 October 2019) https://www.crccare.com/about/the-company/partners
101
UNSW, AI & Robotics Research (accessed 9 October 2019) http://robolab.cse.unsw.edu.au/
102
UNSW Computer Science and Engineering, Research Activities (accessed 9 October 2019)
https://www.engineering.unsw.edu.au/computer-science-engineering/research/research-activities
103
UTS, Centre for Artificial Intelligence (accessed 9 October 2019)

47
https://www.uts.edu.au/research-and-teaching/our-research/centre-artificial-intelligence

MACHINE LEARNING INNOVATION - DECEMBER 2019

08
Conclusion
This report set out to explore patent families in this sector largest source of innovation in
the patent data for innovation alone. Examples of innovation machine learning, with 69 per
relating to machine learning. in telecommunications include cent (25,319) of patent families
The analysis demonstrates internet network operation, having a Chinese applicant. Patent
ongoing commercial measurement of video applications that originate in
development in this sector, with buffering events, prediction of China stay in China - only 4 per
a four-fold global increase in interference between two or cent of applications from Chinese
patent filings over five years more mobile communication applicants are filed internationally.
from 2012−16. Ninety-two per networks and risk assessment The largest applicant is the State
cent of patent families relating for downloading and using Grid, which has filed 1018 patent
to machine learning remain mobile applications. families related to machine
in an active state, indicating learning since 2012.
Image and video analysis is
ongoing commercial value of
the most important machine Following China, the United States,
these patent applications.
learning core capability, South Korea and Japan are the
We found 36,740 machine occurring in 36 per cent second, third and fourth top patent
learning patent families filed of patent families, and the destinations and countries of origin,
since 2012, and analysed foundational technologies of respectively, for patent applications
inventions by machine supervised or semi-supervised relating to machine learning. The
learning techniques and learning techniques are the data shows some technological
their applications. Real world most popular approach, specialisation by country; for
applications of machine accounting for 38 per cent of example the United States is
learning show where machine all patent filings in machine particularly prominent in methods
learning is having an impact learning since 2012. of speech and text analysis.
in current industries. The
China is both the largest patent Australia is an important
telecommunications sector
filing destination with 73 per cent international market in this
leads development of real
(26,758) of machine learning sector, ranking tenth in the
world applications of machine
inventions filed here, and the world as a patent destination for
learning, with 70 per cent of all

48
 4x
PATENT FILINGS
GLOBAL
INCREASE
FROM 2012-16

machine learning innovation. jurisdictions by definition and As well as this, the report
Australia ranks seventeenth interpretation of subject matter includes contributions from
worldwide as a source of eligibility. The WIPO report experts in AI, data, IP, policy,
innovation in machine learning, helpfully sets out an analysis and innovation, adding context
with 95 per cent of Australian of subject matter eligibility to the information revealed from
patent applications relating in the United States, Europe, the patent data, and addressing
to the application of machine Japan, China and Korea, and the issues such as existing and
learning methods. CSIRO is the definition of patentable subject potential uses and impact of AI
leading Australian patent filer in matter in Australia involves technology, legal and regulatory
machine learning. similar considerations. While this questions, data protection and
legal hurdle might be expected to ethical concerns.
Recent publications by WIPO
contribute to patent applications
and the UKIPO give a broader This patent analytics report
lapsing before grant, the data in
overview of technology trends on machine learning shows a
this study shows that only eight
across the whole field of AI. The sector with strong growth and
per cent of machine learning
WIPO report emphasises the potential to boost Australian
patent families are not in an
dominance of machine learning innovation and leverage global
active state, which indicates
techniques, and the major role of investment and growth for
a strong global interest in
China, and also shows that patent the benefit of the broader
pursuing patent protection – and
filings are becoming increasingly Australian economy. This report
commercialisation - in this sector.
important in this sector relative to is a catalyst to consider the
scientific publications. The WIPO report also includes future impact of AI and inform
further analysis on market policy and decisions to ensure
In the field of machine learning,
trends related to AI, including that Australians benefit from
it is also important to be aware of
company mergers and machine learning.
difficulties in patenting computer
acquisitions, patent oppositions,
algorithms. Patentability
patent litigation, public funding,
of computer-implemented
and investment in start ups.
inventions varies across

MACHINE LEARNING INNOVATION - DECEMBER 2019 49

09 Appendix A: Definitions

Patent, applications and directly in the countries of interest, IPC or CPC systems. The CPC began
publications without using the PCT system. in 2013 and provides significantly
more depth to the hierarchy of the
A patent is a right that is granted
Patent families IPC. For more information on the
for any device, substance, method
Applications with the same priority, coverage of the CPC, see the CPC
or process that is new, inventive
but filed in different jurisdictions, Annual Report 2017–18.
and useful. Australian patent rights
are legally enforceable and give are known as patent families.
Patent families enable us to analyse Patent status determination
the owner, or patentee, exclusive
rights to commercially exploit the inventive activity regardless of A patent application is considered
invention in Australia for a period the number of countries in which as ‘being sought or in force’ when
of up to 20 years. In this report, an protection is sought. Patent families it has not lapsed (due to expiry or
application refers to a single patent are used in analytics to represent non-payment of renewal fees), been
filing. A patent application is usually a single invention. We determine revoked or withdrawn. A family
published within 18 months of its patent families based on DOCDB is designated as ‘being sought or
earliest filing date (also known as database definition, with a unique in force’ if it contains at least one
the priority date). We consider that family ID for patents that share application that is ‘being sought or
the priority date is most relevant to a common priority document. in force’. The PATSTAT database
our analysis as it is the closest date The number of patent families is used to define the legal events
to that when the invention occurred. is typically used as a metric. ‘Application Discontinuation’ and
There are some exceptions when ‘IP Right Cessation’. These are
There are two major routes for reporting individual applications, considered to make an application
filing a patent application: the as each application represents a lapsed, expired or withdrawn.
international route and direct filing. legal right in an individual country. Applications without these events
The international route involves filing When analysing applicants, related are considered as ‘being sought or
a PCT application, which establishes commercial entities are grouped by in force’.
a filing date in all 152 contracting a singly, harmonised name. When
states. Subsequent prosecution at individual publication numbers
national patent offices, referred to as are quoted, we have chosen a
national phase entry, is made at the representative publication from the
discretion of the applicant. A patent patent family, typically US or WO
can only be enforced once it has been English language documents.
granted and a PCT application must
enter the national phase in each Classification
country to proceed toward grant. Patents are hierarchically classified
Alternatively, applications can be filed by technology into the hierarchical

50
Appendix A: Definitions

Country of origin applicant name with several country information for the same
entries for country code. A reason, even for local applications.
Many records in the PATSTAT
count of the total number The first application authority is
database (around 50 per cent) do
of applications applied for therefore generally only used as
not have any associated country
with each country code is a proxy for origin in the case of
code information for applicants/
patent families that are filed to
inventors. In previous patent used to determine the most
one jurisdiction only, increasing
analytics studies, null data has representative country and that is
the probability that the country is
generally been excluded from the applied. This is an independently
correctly assigned.
country of origin analysis. developed technique analogous
to the technique detailed by the
In this patent analytics report, we
have used other data available European Commission.
in PATSTAT to provide additional 3. If the country code for the
insights for patent families that standardised name is ‘null’
do not have applicant origin data or where no applicant details
available. This method of country
are recorded for the patent
code assignment leverages all
application, then the first
data from the entire PATSTAT
application authority is used.
database, and not only country
This is the country the earliest
codes information from a subset
of the data extracted for a specific priority document was filed
technology area. into for the patent family and it
is used as a country of origin.
The country of origin of a patent This authority is the first by date
family is assigned using a three- where the authority is a country
step process.
and not an international body.
1. Where country code is available For international applications,
for applicant data, this is used the receiving office is used
as the country of origin. where possible.

2. If the country code is ‘null’ and It is important to note that as patent

families are used, most families
the applicant name is associated
that are filed internationally have
with a standardised or cleaned
reliable country code data, either
version of their name (e.g. the
through WIPO, EPO, USPTO or other
PATSTAT standardised PSN
large patent offices. Applicants that
name), there may be several often file internationally will have
versions of an associated standardised name-associated

MACHINE LEARNING INNOVATION - DECEMBER 2019 51

 Appendix B: Search Strategy

Searching patent information to identify relevant records Classification Symbols (where a truncation is given, all
for analysis requires a stepped approach to identify elements in the hierarchy beginning with that stem are
broad categories of relevance, and then specific records included).
within them that meet the technology brief.
1. IPC Symbols:
The following details outline the search and analysis G06N3/02, G06N3/04, G06N3/06%, G06N3/08,
process conducted. G06T1/40

2. CPC Symbols:
Data extraction and analysis
G06N3/02, G06N3/04%, G06N3/06%, G06N3/08%,
We used five phases of data extraction and analysis. G06N3/10, G06N3/105, G06N7/046, G06N 99/005,
G06T2207/20081, G06T2207/20084, G06T3/4046,
• Phase 1: Development of a search strategy (below).
G06T9/002, G05B13/027, G05B13/0285, G05B13/029,
• Phase 2: Data mining using the database PATSTAT
G10L17/18, G01N2201/1296, G01N29/4481,
2018 Autumn edition. The unique DOCDB family
G01N33/0034, G01S7/417, G06F11/1476, G06F11/2263,
members relating to machine learning were identified
G06F2207/4824, G06N7/046, G11B20/10518,
and used as the basis of the analysis.
G10H2250/311, H02P21/0014, H02P23/0018,
• Phase 3: Data cleaning, focusing on removing errors H03H2017/0208, H03H2222/04, H04L2012/5686,
and harmonising applicant names and ensuring the H04L2025/03464, H04L2025/03554, H04L25/0254,
return of correct records. This was done in Derwent H04L25/03165, H04L45/08, H04N21/4662,
Data Analyser version 7. H04N21/4663, H04N21/4665, H04N21/4666,
• Phase 4: Data categorisation according to the H04Q2213/13343, H04Q2213/343, B60G2600/1878,
technological focus of the patent families. The B23K31/006, B29C2945/76979, B29C66/965,
technological focus was determined using the CPC Y10S128/925, F02D41/1405, F05B2270/709,
and IPC symbols. F16H2061/0084, F05D2270/709, G10K2210/3038,
G10L25/30, G10L15/16
• Phase 5: Data analysis using Tableau 2018.3 for
calculation and visual presentation of patent metrics. Keywords:

Search strategy An exact keyword search involves tokenising the field in

question and then finding the exact word in the list. In
The search was limited to patent applications that had
the majority of cases unlimited left and right truncation
an earliest priority date between 201218 and a CPC
was used. This has been summarised for clarity, but it
symbol, IPC symbol, keywords (in the title or abstract of
means that the words in the order provided were found
the patent) or a combination of both symbol and keyword
anywhere in the field in question in exactly the order
as listed here. (“%” indicates unlimited truncation.)
shown with any number of characters on either side,

52
Appendix B: Search Strategy

including prefixes or suffixes. An underscore represents boosting, xgboost, adaboost, adaptive boost, adaptive_
exactly one wildcard character. boost, rankboost, boosting algorithm, Bootstrap
aggregat, stacked generali_ation, stacked_generali_
1. Searched exactly:
ation, logistic regression, stochastic gradient descent,
convnet, convnets, LSTM, LPboost, Totalboost,
latent semantic analysis, latent dirichlet allocation,
brownboost, madaboost, Logitboost
meta learning, Q-learning, temporal difference
2. Searched with unlimited left and right truncation: learning, learning automata, learning_automata,
neural net, multi_layer perceptron, multi_layer extreme_learning_machine, extreme learning
perceptron, machine learning, training algorithm, machine, logic learning machine, single_linkage
deep learning, deep belief net, deep_belief net, clustering, single linkage clustering, conceptual
generative adversarial net, long short term memory, clustering, multilayer perceptron, deepbelief
long_short_term memory, multi label classif, multi_ net, longshort_term memory, long_shortterm
label classif, random forest, reinforcement learning, memory, longshortterm memory, multilabel
semi supervised learning, semi_supervised learning, classif, semisupervised learning, autoencoder,
supervised learning, unsupervised learning, similarity rulebased learning, swarm behavior, k-nearest
learning, dictionary learning, Hebbian learning, neighbor, mixturemodel, selforganising map, self
feature learning, auto encoder, auto_encoder, organising map, selforganizing map, self organizing
support vector regression, support vector machine, map, Instancebased learning, multitask learning,
rule learning, rule based learning, rule_based adaptiveboost, learningautomata, extremelearning_
learning, decision tree, classification tree, regression machine, extreme_learningmachine,
tree, swarm intelligence, swarm optimisation, swarm extremelearningmachine, extremelearning machine,
optimization, swarm behavio_r, kernel learning, extreme learningmachine, singlelinkage clustering
k-nearest neighbo_r, k-mean, hierarchical cluster,
3. Searched in a combination of exact (without
naïve Bayes, naive Bayes, Bayesian net, Bayesian
% wildcards) and with unlimited left and right
learning, mixture model, mixture_model, Blind
truncation, as listed:
signal separation, independent component analysis,
(inference & %learn%), (inference & %train%)
self_organising map, self organising map, Markov%
decision process, natural language processing,
chatbot, Virtual assistant, probabilistic graphical
model, Instance_based learning, Instance based
learning, multi_task learning, multi task learning,
statistical relational learning, transfer learning,
ensemble learning, gradient boosting, gradient tree

MACHINE LEARNING INNOVATION - DECEMBER 2019 53

 Appendix C: Technology Analysis

Patent families relating to machine learning were first categorised into two
groups, machine learning techniques and machine learning applications,
using IPC/CPC symbols and keywords. This classification was done in
Tableau Desktop.

Multiple classification symbols may be present for a single patent family

and in categorising patent families, overlap where families could be placed
in multiple sub-categorises were noted. Sub-categories were prioritised as
per the order of the tables below to ensure a patent family appears only in
one sub-category. Prioritisation was not done between machine learning
foundational technologies and machine learning applications as one patent
family could inherently discloses both the foundational technologies and
their applications.

Machine learning foundational

technologies categorisation
The machine learning techniques category includes patent families that fall
into algorithms and architecture.

Algorithms
Algorithms further includes supervised and semi-supervised learning,
unsupervised learning, reinforcement learning, niche learning techniques and
general techniques sub-categories as per the priority order shown below.

PRIORITY SUB-CATEGORY COMPONENTS IPC CPC KEYWORDS

Reinforcement [Dynamic OR (dynamic programming, markov

1
learning Programming] decision process%)

[Learning (learning_automat%)
Automata]

[Reinforcement OR ((reinforce% AND (neural_

Neural network%)), deep reinforce%,
Networks] deepmind, deep q network%, DQN,
DDPG, deep deterministic policy
gradient%)

[Other OR (reinforce% learn%, q_learn%,

Reinforcement sarsa) AND NOT ([Dynamic
learning] Programming] OR [Learning
Automata] OR [Reinforcement
Neural Networks])

54
Appendix C: Technology Analysis

PRIORITY SUB-CATEGORY COMPONENTS IPC CPC KEYWORDS

2 Supervised [Boosting] OR (boosting, meta_algorithm,

semi-supervised adaboost, lpboost, totalboost,
learning brownboost, xgboost, logitboost,
adaptive_boost, rankboost)

[Classification OR (decision tree, random

and Regression] forest, classif% tree, svm,
support vector machine%, k_NN,
nearest_neighbo[, u]r, similarity
learning, naïve bayes, bayes%
network%, regression, linear
discriminant analysis, gaussian
process%, bayes% learn, binary_
classification%, multi_label_
classification, kernel, classification
model%, classifier%)

[Supervised (supervi[s,z]% AND (dictionary

Dictionary] learning))

[Supervised G06N3/084 OR (supervised neural net,

Neural convolution% neural net, inferred
Networks] function, convnet, CNN, error
signal, label% training data,
label% data, supervisory signal%,
supervised teach%, multi-layer
perceptron%, back-prop%, radial
basis function%, RBF)

[Supervised ((supervis%, semi_supervis%,

Learning- weekly_supervis%) AND NOT
General] ([Supervised Neural Networks] OR
[Classification and Regression]
OR [Boosting] OR [Supervised
Dictionary]))

3 Unsupervised [Blind Signal OR (blind signal separation%,

learning Separation] principle component analysis,
independent component analysis,
matrix factori[s,z]ation%)

[Cluster OR (cluster%, hierarchical, k_

Analysis] mean%, mixture model%, dbscan,
optics algorithm%)

[Dictionary (dictionary learning) AND NOT

Learning] [Supervised Dictionary]

MACHINE LEARNING INNOVATION - DECEMBER 2019 55

 Appendix C: Technology Analysis

PRIORITY SUB-CATEGORY COMPONENTS IPC CPC KEYWORDS

[Unsupervised G06N3/088 OR (non_supervised neural_net,

Neural unsupervised neural_net,
Networks] auto_encoder, unlabel_ed data,
deep belief net, hebbian learning,
self_organi[s,z]ing map, generative
adversarial network, competitive_
learn)

4 Niche learning [Transfer OR (transfer learn, inductive learn,

techniques Learning] inductive transfer, transductive
transfer)

[Rule-Based OR (rule_based_machine_learning,
Learning] RBML)

[Relational (relation% learn%)

Learning]

[Particle Swarm (particle swarm, swarm

Optimisation] intelligence, swarm optimi%,
swarm behavi%)

[Online learning] OR (online learn%, online machine

learn%)

[Graphical (graphical model%)

Model]

[Genetic G06N3/12% G06N3/086, (genetic algorithm%)

Algorithms] G06N3/12

[Gaussian (mixture model%)

Mixture Model]

[Feature (feature learn%)

Learning]

[Extreme OR (extreme learn%, ELM)

learning]

[Ensemble (ensemble learn%)

learning]

[BCM Rule OR (bcm rule, ibcm rule, bcm

learning] learning%)

[Ant Colony (ant_colony)

Optimisation]

56
Appendix C: Technology Analysis

PRIORITY SUB-CATEGORY COMPONENTS IPC CPC KEYWORDS

[Learning Paradigms] = [Supervised Learning] or [Unsupervised Learning] or [Reinforcement Learning] or [Niche Learning]

8 General [Machine- G06N99%, G06N3/10%, (OR (machine_learn%, deep

techniques Learning- G06N3/02%, G06N3/08%, neural net%, deep learn%, deep
general] G06N3/10%, G01N33/0034%, belief net%, DNN, neural_net%.
G06N3/08% H04L45/08%, Learning_machine%, training
G06N3/105%, algorithm%) AND NOT ([Learning
G06F11/1476%, Paradigms] OR [NN Architecture])
G06F11/2263%,
G06F2207/4824%,
H04L2012/5686%,
H04L45/08%,
H04Q2213/13343%,
H04Q2213/343%,
B60G2600/1878%,
B23K31/006%

[Gradient OR (stochastic gradient descent%,

Descent] regulari[s,z]ation%)

[General techniques] = ([Machine-Learning-general] OR [Gradient Descent]) AND NOT [Applications]

% is a wildcard for zero or more characters

_ is wildcard for zero or one character

Architecture
The architecture sub-category further includes neural network architecture, architecture optimisation and physical
realisations of neural networks as per the priority order shown below.

PRIORITY SUB-CATEGORY IPC CPC KEYWORDS

5 Architecture optimisation G06N3/082

6 Neural network architecture G06N3/04% G06N3/04%

7 Physical realisations G06N3/06% G06N3/06% (OR (FPGA, fixed point, hardware

implement%, hardware reali[s,z]) AND
(neural net%)) OR (graphic% process%
unit%)

% is a wildcard for zero or more characters

_ is wildcard for zero or one character

MACHINE LEARNING INNOVATION - DECEMBER 2019 57

 Appendix C: Technology Analysis

Machine learning applications categorisation

The machine learning applications are categorised by core capabilities and/or real world applications.

Core capabilities
The machine learning applications category includes patent families that fall into speech and text analysis, image
and video analysis, control, optimisation and automation, signal analysis and data analytics and processing as per the
priority order shown below.

PRIORITY SUB-CATEGORY IPC CPC KEYWORDS

1 Speech and text G10L%, G06F17/2 G10L%, G06F17/2 OR (natural language process%, language
analysis model%, natural language understand%,
NLP, NLU, word embedding, tokeniz%,
semantic pars%, entity extraction%, entity
name extraction%, entity resolution%,
part_of_speech, POS, LSTM, long_
short_term_memory, attention model%,
pronunciat%, speech%, audio process%,
chat_bot, virtual assist%)

2 Image and video G06T%, G06K9% G06T%, G06K9% OR (image analy%, image classif%, image
analysis process%, image compar%, image mat%,
object recognition, object classif%, object
match%, target recognition%, target
classif%, target match%, spectral imag%,
super resolution%, depth perception%,
pose estimate%, captioning, adagrad%,
adam gradient%, (OR (audio, media) AND
OR(feature extract%, feature analy%,
feature classif%)))

3 Control, B25J9%, B25J9%, B21B37%, OR (automatic control%, control system%,

optimisation and B21B37%, B29C64/393%, control method%)
automation B29C64/393%, B33Y50/02%, G05B%,
B33Y50/02%, G05D%, F01N2900/0402%,
G05B%, G05D%, F02D41/2438%,
H03L%, H02P%, F02D41/2441%,
F24F11% F02D41/1405%,
F03D7/046%,
F05B2270/709%, H03L%,
H02P%, F24F11%,
B29C2945/76979%,
B29C66/965%,
F16H2061/0084%

58
Appendix C: Technology Analysis

PRIORITY SUB-CATEGORY IPC CPC KEYWORDS

4 Signal analysis A61B5%, A61B8%, A61B5%, A61B8%, OR (signal% AND OR (process%,

G01N21%, G01N21%, G01N22%, analy[s,z]%))
G01N22%, G01N27/90%, G01R%,
G01N27/90%, G01N29/4481%,
G01R% G01N33/0034%,
G10H2250/311%,
G01S7/417%,
H04L25/03165%,
H04L25/0254%, H04L2025/
03464%,H04L2025/03554%,
G11B20/10518%,
H03H2017/0208%,
H03H2222/04%,
G10K2210/3038%

5 Data analytics and G06Q%, G06F3%, G06Q%, G06F3%, G06F5%, OR (data mining, data analy%, analy%
processing G06F5%, G06F7%, G06F7%, G06F16%, data, cognitive, data filter%, filter% data)
G06F16%, G06F17/14%, G06F17/15%,
G06F17/14%, G06F17/16%, G06F17/30%,
G06F17/15%, G06F17/40%, G06F2216%,
G06F17/16%, Y10S706%, H04N21/466%
G06F17/30%,
G06F17/40%,
H04N21/466%

% is a wildcard for zero or more characters

_ is wildcard for zero or one character

Real world applications

The machine learning applications category includes patent families that fall into predictive maintenance,
telecommunication, agriculture, energy, environment and mining, security and military, retail, health care, industry
and manufacturing, transport and logistics and, business and finance as per the priority order shown below.

PRIORITY SUB-CATEGORY IPC CPC KEYWORDS

1 predictive G05B23/0283% G05B23/0283% OR ((predict% AND maintenance), (OR

maintenance (fault%, wear%, tear%, acoustic%,
failure%, defect) AND OR (diagnos%,
detect%, predict%, identif%, calculate%,
monitor%)))

OR (diagnos% and current%, damage%

and identif%)

MACHINE LEARNING INNOVATION - DECEMBER 2019 59

 Appendix C: Technology Analysis

PRIORITY SUB-CATEGORY IPC CPC KEYWORDS

2 telecommunication H04B%, H04H%, H04B%, H04H%, H04J%, OR (telecom%, celllular network%,

H04J%, H04K%, H04K%, H04L%, H04M%, mobile network%, computer network%,
H04L%, H04M%, H04N%, H04Q%, H04R%, internet security, data security, LTE,
H04N%, H04Q%, H04S%, H04T%, H04W% LTE-A, lte_advance%, malware%, spam%,
H04R%, H04S%, IOT%, social network%, communication
H04T%, H04W% network%, positioning, cloud service%,
network system%, server system%, audio
coding, video coding, network resource%,
communications network%)

OR (webpage, coaxial_cable, communicaiton,

mimo, multiple_input_multiple_output,
optic__fiber, optic__fibre, radar)

3 agriculture A01% A01%, G05D2201/0201%, OR (plant cultivat%, farming, irrigate%,

Y02A40/1%, agriculture%, crop, grain, live_stock%,
Y02A40/2%, Y02A40%, vegetation%, honey) AND NOT (honeycomb%)
G06T2207/30188%
OR (fruit, maturity, milk, cotton, plant,
grapefruit, beef, brooder, meat, garlic, wheat,
tobacco, animal feed, fish, sorghum, grain,
algae, tea, microalgae, peanut%, protein,
food, ferment, gelatin, liquor, vintage,
drink, hogwash, gutter, oil, pork, apples,
shrimp, chicken, sausage%, citrus, rice,
maize, chlorophyll, eggs, baijiu, cooking
apparatus, straw, sugar, wine, pear tree,
cooking appliance, red date, cyclotella, baking
method, cooking utensil, pest, vaporariorum,
quarantine, larvae, pesticide)

4 energy F01%, F02%, F01%, F02%, F03%, F04%, OR (voltage%, power consumption%,
F03%, F04%, F15%, F17%, H01%, H02%, power distribut%, power generat%,
F15%, F17%, H03%, Y10S706/915%, power transmit%, electric%,
H01%, H02%, Y10S706/907%, transformer%, power station%, power
H03%, Y10S60/00%, G06Q50/06%, plant%, harmonic emission%, partial
G06Q50/06%, G01R21% discharge%, battery, multimeter%,
G01R21% power grid%, energy management%,
alternating_current%, solar, wind power,
energy saving%, microgrid|transmission
line%,transmission wire%, wind farm%,
photovoltaic%, line_to_ground%) OR (
OR(power,energy) AND OR(load predict%,
output predict%, load forecast%))

OR (gyro, diode, fpga, circuit board,

semiconductor, grounding grid,
electromagnetic, boiler, analog circuit,
analogue circuit, coaxial cable, integrated
circuit, electronic circuit, line loss, energy
billing, 10kv line, resistive, energy network,

60
power system, microwave filter)
Appendix C: Technology Analysis

PRIORITY SUB-CATEGORY IPC CPC KEYWORDS

5 environment and E21%, G01W%, E21%, G05D2201/021%, OR (excavat%, geology%, seismology%,

mining G01V%, Y10S706/928%, weather%, meteorology, seismic%,
G01C11%, Y10S706/929%, earthquake, (OR (coal%, opencast%,
G01C13%, Y10S706/93%, Y10S37/00%, mineral%, fossil%, oil, gas, LNG, gold) AND
G01C15% G01W%, G06T2207/30192%, (mine%, mining, drill%, mill%)), (earth% and
G01V%, G01C11%, drill%))
G01C13%, G01C15%
OR (oil_gas_water, landslip, forest fire, soil,
oxygen, housing, biomass furnace, fire alarm,
garbage calorific, wind, typhoon, thunder,
seabed, sea surface, fire flame, environment,
water level, water tank, algal bloom, fire
disaster, fire-hazard, trash content, gask
leak, water quality, smog, rock, well line,
gis, atmospheric, ground feature, soil, river,
remote sensing, coal, forest, landmark,
aerial photograph, areally photograph,
oil_producing well, oil well, oil overflow,
air_quality, geomagnetic, plankton, pressure
vessel, terrain, aerial scene, polari[sz]ed
sar, sar image, sediment, bird population,
pollutant, heavy metal, natural gas, ore
body, water analysis, flame, fire AND (detect,
chemical) AND (contamination))

6 security and OR (forgery, copy protection, public

military sentiment, public monitor, intrusion,
alarming, surveillance, counterfeit, ground
target, camouflaged, trojan, detecting
attack, fire control, authentication,
spoofing, gun detection, war, crime,
missile, submarine, violent behaviou?r, sql
injection, threat analysis)
OR (face, facial, finger_print, iris, voiceprint,
speaker recognition, speaker age,
handwriting, handwritten, hand recognition,
palm_print, palm detection, hand_print)

7 retail G06Q30%, G06Q30%, G06Q50/28% OR (product% AND OR (recommend%,

G06Q50/28% identif%, personali%), demand predict%,
predict% demand, customer engage%,
customer satisf%, personali% market%,
personali% adverti%, telemarket%, water
meter%, amazon_go)
OR (e_commerce, electronic_commerce,
e_shopping, electronic_shopping,
internet_commerce, internet_shopping,
web_commerce, web_shopping, shopping
mall, clothing, textile, fabric, clothes)

MACHINE LEARNING INNOVATION - DECEMBER 2019 61

 Appendix C: Technology Analysis

PRIORITY SUB-CATEGORY IPC CPC KEYWORDS

8 health care A61%, G16B%, A61%, G16B%, OR (medicin%, medical%, biology, biological,
G16H%, G16H%, G06Q50/24%, patient%, health_care, prosthesis, neural
G06Q50/24%, G06F19%, G01N2800%, oscillator%, tumour%, tumor%, disease%,
G06F19%, Y10S706/924%, Y10S128%, surgery, surgical, clinical, cancer, MRI,
C12M1% C12M1% imag%, genetic toxicity, IVF, cardiac%,
peptide, influenza, bacteria)
OR (urine, endoscopic, protein process,
blood cell, visual cortex, x-ray, emotion,
cleavage cell, mammogram, vivo,
electromyography, gastroscope, eye fundus,
developmental disorder, nuclear magnetic
resonance, magnetic resonance imag%,
tongue, ecg, stem cell, marrow, biomarker,
patholog%, cell outline, cell boundary, vein,
vascular, bone, cilia, pelvic, neurotox%,
thyroid, health management, lung, plasmid,
retina, blood, cerebral, vessel, artery, fmri,
mammary, adipose, abdominal, lesion,
pulmonary, cervical, tooth, humoral, breast,
tissue, fall, tumble AND (detection, monitor,
brain machine interface, brain computer
interface, encephalogram, eeg))

9 industry and B01%, B02%, B01%, B02%, B03%, B04%, (manufactur%, process % tech%,
manufacturing B03%, B04%, B05%, B06%, B07%, galvaniz%, aluminium cast%, copper
B05%, B06%, B08%, B09%, B21%, B22%, concentr%)
B07%, B08%, B23%, B24%, B25%, B26%,
B09%, B21%, B27%, B28%, B29%, B30%,
B22%, B23%, B31%, B32%, B33%, B65%,
B24%, B25%, B66%, B67%, B68%, C02%,
B26%, B27%, C03%, C04%, C05%, C10%,
B28%, B29%, C13%, C21%, C22%, C23%,
B30%, B31%, D06%, D21%, F01%, F02%,
B32%, B33%, F03%, F04%, F05%, F16%,
B65%, B66%, F22%, F23%, F24%, F25%,
B67%, B68%, F27%, G05D2201/0201%,
C02%, C03%, G05D2201/0216%,
C04%, C05%, G06Q50/04
C10%, C13%,
C21%, C22%,
C23%, D06%,
D21%, F01%,
F02%, F03%,
F04%, F05%,
G06Q50/04

62
Appendix C: Technology Analysis

PRIORITY SUB-CATEGORY IPC CPC KEYWORDS

10 transport and B60%, B61%, B60%, B61%, B62%, OR (autonomous vehicle%, autonomous
logistics B62%, B63%, B63%, B64%, G05D1%, car%, driverless, driver assist%, traffic%,
B64%, G05D1%, G05D2201/0212%, supply chain%, automatic driving, auto
G08G%, G05D2201/0213%, driving, autonomous driving)
G06Q10/08%, G06Q2240/00%, OR (driver, pedestrian, car, road sign,
G06Q50/28%, G08G%, Y10S706/905%, license plate, steel rail, taxi, postal,
G06Q50/30% G06Q10/08%, G06Q50/28%, transport, driving, truck, plate number,
G06Q50/30%, vehicle color, chassis, ship, vehicle,
G06T2207/30181% vehicular, high_speed, rail, railway, road,
aerocraft, airplane, aeroplane, aircraft,
rotorcraft, navigation, automobile, flight
control, uav, locomotive, helicopter,
cockpit, bus, busses, tire, aerospace, lane
AND (recogni[sz]%, detect%, recognit%,
elevator, escalator) AND (visitor, people,
passenger))
OR (satellite, spaceflight, spacecraft)
OR ((OR (earth,planet) AND OR (observ%,
monitor%)), (space_station%)

11 business and G06Q10/02%, G06Q10/02%, G06Q10/04%, OR (trading,robo advisor%, finance, financial,

finance G06Q10/04%, G06Q10/06%, G06Q10/10%, transaction%, banking, audit,
G06Q10/06%, G06Q20%, G06Q40%, (OR (risk%,credit%,portfolio, workforce)
G06Q10/10%, Y10S706/925%, AND (analysis, assessment%, approv%,
G06Q20%, Y10S706/926%, management%, planning)),
G06Q40%, (OR (currency,money,stock%) AND
G09B%, G09B%, G06Q50/20%, (exchang%, laundering))
G06Q50/20, Y10S706/927%,
OR (advertisement filtering, movie
A63% A63% expectation, house property, merchandise,
insurance, stock news, investor)

OR (education%, teaching%, guided

learning%)

OR (pronunciation, student, classroom)

OR (video game%, board game%, sport%,

ball game%, skate%, skating%, card game%,
indoor game%, outdoor game%, bowling,
racing, riding)
OR (sport, augmented reality, head_
mounted_ display, virtual reality, badminton,
basketball, online game, computer game)”

MACHINE LEARNING INNOVATION - DECEMBER 2019 63

 Abbreviations

ACS Australian Computer Society

AI Artificial intelligence
BoF Bag of features
CAI Centre for Artificial Intelligence
CNN Convolutional neural networks
CSIRO Commonwealth Scientific and Industrial Research Organisation
CSPG China Southern Power Grid
CT Computed tomography
DFKI Deutsches Forschungszentrum für Künstliche Intelligenz
(German Research Center for Artificial Intelligence)
FoG Freezing of Gait
GAN Generative adversarial network
GIS Gas insulated switchgear
IBM International Business Machines
IP Intellectual property
IPO Intellectual Property Office (UK)
LSTM Long-short-term memory
MIT Massachusetts Institute of Technology
MRI Magnetic resonance imaging
NLP Natural language processing
NICTA National ICT Australia
NTT Nippon Telegraph and Telephone Corporation
ODN Object detection network
PCT Patent Cooperation Treaty
RNN Recurrent neural network
RPN Region proposal network
SMILE Statistical Machine Intelligence and LEarning (UESTC)
State Grid State Grid Corporation of China
UESTC University of Electronic Science and Technology of China
UKIPO United Kingdom Intellectual Property Office
WIPO World Intellectual Property Office

64
 About ACS

ACS is the professional

association for Australia’s
Information and Communication
Technology (ICT) sector. More
than 44,000 ACS members
work in business, education,
government and the community.
ACS exists to create the
environment and provide the
opportunities for members
and partners to succeed. ACS
strives for ICT professionals
to be recognised as drivers
of innovation in our society,
relevant across all sectors, and
to promote the formulation of
effective policies on ICT and
related matters.
Visit acs.org.au for
more information.

MACHINE LEARNING INNOVATION - DECEMBER 2019 65

66
ACS
International Tower One
Level 27, 100 Barangaroo Avenue
Sydney NSW 2000
P: 02 9299 3666
F: 02 9299 3997
E: [email protected]
W: acs.org.au