Think Big! The Need for Patent Rights in the Era of Big Data and Machine Learning

By: Hyunjong Ryan Jin*

Download a PDF version of this article here

Introduction

With AlphaGo’s triumph over the 9-dan Go professional Lee
Sedol in March 2016, Google’s DeepMind team conquered the last remaining
milestone in board game artificial intelligence.[1] Just nineteen years after IBM Deep
Blue’s victory over the Russian chess grandmaster Garry Kasparov,[2] Google’s success
exceeded expert predictions by decades.[3]

AlphaGo demonstrated how machine learning algorithms could
enable processing of vast amounts of data. Played out on a 19 by 19 grid, the
number of possible configurations on a Go board is astronomical.[4] With near-infinite
number of potential moves, conventional brute-force comparison of all possible
outcomes is not feasible.[5] To compete with professional
level human Go players, the gaming artificial intelligence requires a more
sophisticated approach than the algorithms employed for chess — machine
learning. The underlying science and implementation of machine learning was
described in a Nature article two months prior to AlphaGo’s match with Lee. In
the article, the Google team described how a method called “deep neural
networks” decides between the insurmountable number of possible moves in Go.[6] The AlphaGo model
was built by reinforcement learning from a database consisting of over thirty
million moves of world-class Go players.[7] This allowed the
algorithm to optimize the search space of potential moves, therefore reducing
the required calculations to determine the next move.[8] In other words, the
algorithm mimics human intuition based on the “experience” it gained from the
database “fed” into the algorithm, which drastically increases computational
efficiency by eliminating moves not worth subsequent consideration. This allows
the algorithm to devote computational resources towards the outcomes of
“worthwhile” moves.

The
advent of such powerful analytical tools, capable of mimicking human intuition
alongside massive computation power, opens endless possibilities—early stage
cancer detection[9], accurate weather forecasting,[10] prediction of corporate
bankruptcies,[11] natural event detection,[12] and even prediction of
elections.[13] For information technology (“IT”)
corporations, investment in such technology is no longer an option, but a
necessity. The question that this Note addresses is whether the current state
of intellectual property law is adequate to harness the societal benefits that
we hope to enjoy through the advances in machine learning. In particular, are
patents necessary in the age of big data? And if they are, how should we apply
patent protection in the field of big data and machine learning?

Part I
of this Note examines the need for intellectual property rights in machine
learning and identifies the methods by which such protection may be achieved.
The differences between trade secret, copyright and patent protection in
software are discussed, followed by the scope of protection offered by each
means. This background provides the basis to discuss the effectiveness of each
method in the context of machine learning and big data innovations.

Part II
discusses the basics of the underlying engineering principle of machine
learning and demonstrates how the different types of intellectual property
protection may apply. Innovators may protect their contributions in machine
learning by defending three areas—(1)
the vast amount of data required to train the machine learning algorithm, (2)
innovations in the algorithms itself including advanced mathematical models and
faster computational methods, and (3) the resulting machine learning model and
the output data sets. Likewise, there are three distinct methods of protecting these intellectual properties: patents,
copyright, and secrecy.[14] This Note discusses the
effectiveness of each method of intellectual property protection with three
principles of machine learning innovation in mind: facilitating data sharing,
avoiding barriers to entry from data network effects, and providing incentives
to address the key technological challenges of machine learning. This Notes
proposes that patents on computational methods adequately balance the concern
of patent monopoly and promoting innovation, hence should be the primary means
of intellectual property protection in machine learning.

Part
III then visits the legal doctrine of patentable subject matter starting with
the United States Supreme Court’s Alice
decision. While Alice imposed a high
bar for software patents, the post-Alice
Federal Circuit decisions such as Enfish,
Bascom, and McRO suggest that certain types of software inventions are still
patentable. Specifically, this section will discuss the modern framework
pertinent to subject matter analysis: (1) inventions that are directed to
improvements of computer functionality rather than an abstract idea, (2)
inventions that contain an inventive concept, and (3) inventions that do not
improperly preempt other solutions. The Note will apply this framework to
innovations in machine learning.

The
Note proposes that patents for computational methods balance the need for
intellectual property protection while permitting data sharing, paving the
pathway for promoting innovation in machine learning. The Note further argues
that machine learning algorithms are within patentable subject matter under 35
U.S.C. §101.

I. Need for Intellectual Property Rights in Machine Learning

“He who receives an idea from me, receives
instruction himself without lessening mine; as he who lights his taper at mine,
receives light without darkening me.”

–
Thomas Jefferson

“I’m going to destroy Android, because it’s a
stolen product. I’m willing to go thermonuclear war on this. They are scared to
death, because they know they are guilty.”

– Steve
Jobs

The two
quotes above demonstrate the conflicting views on protecting intangible ideas
with intellectual property law. Thomas Jefferson implied that the free
circulation of inventive ideas and thoughts would not dampen the progress of
innovation nor disadvantage innovators. On the other hand, Steve Jobs exhibited
fury over the similarity between the iOS and the Android OS. Why? Was it
because his company was worse off due to the similarity between the two
products? Would Apple have refrained from inventing the iPhone had it known
others would enter the smartphone market?

This
section discusses the motives behind the grant of intellectual property rights
and whether such protection should be extended to machine learning innovations.
Basics of patent law, copyright law, and trade secret are introduced to provide
the analytical tools for subsequent discussion on which type of intellectual
property protection best promotes the socially-beneficial effects of machine
learning.

A. Do We Need Intellectual Property Rights for Machine Learning?

The
primary objectives of intellectual property rights are to encourage innovation
and to provide the public with the benefits of those innovations.[15] In the context of machine learning, it is not clear whether
we need any additional incentives to promote participation in this field.
Machine learning is already a “hot field,” with countless actors in industry
and academia in active pursuit to keep pace.[16] Hence investment incentivizing may
not be a valid justification for granting intellectual property rights in
machine learning. Rather, such protection is crucial to promote competition and
enhance public benefits.

The quality of inferences that may be drawn
from a given data set increases exponentially as the aggregation diversifies,
which is why cross-industry data aggregation will greatly enhance the societal
impact of machine learning.[17] Companies will need to identify new data access points
outside of their own fields to gain access to other data sets to further
diversity their data. Yet the incentive structures of behemoth corporations may
not be well-suited to identify and grow niche markets.[18] It would be up to the smaller,
specialized entities to find the gaps that the larger corporations overlooked
and provide specialized services addressing the needs of that market.
Protective measures that assist newcomers to compete against resource-rich
corporations may provide the essential tools for startups to enter such
markets. Sufficient intellectual property protection may serve as leverage that
startups may use to gain access to data sets in the hands of the Googles and
Apples of the world, thus broadening the range of social benefits from machine
learning.

B. The Basics of Patent Law

“To promote the progress of science and
useful arts, by securing for limited times to authors and inventors the
exclusive right to their respective writings and discoveries” – United
States Constitution, Article I, § 8

The
United States Constitution explicitly authorizes Congress to promote useful
arts by granting inventors the exclusive rights of their discoveries. Such
constitutional rights stems from two distinct bases — (1) a quid pro quo where
the government issues a grant of monopoly in exchange for disclosure to
society, and (2) property rights of the inventor. The purpose for such rights
is explicitly stated in the Constitution—to promote new inventions. The goal is
to prevent second arrivers who have not invested in the creation of the initial
invention from producing competing products and services at a lower price,
undercutting the innovator whose costs are higher for having invested to create
the invention. As an incentive for innovators willing to invest in new, useful
arts, the patent system provides the innovator rights to exclude others from practicing
the invention. Another purpose of such rights is the concept of “
“mining
rights.” Akin to the grant of mining rights to the owner in efforts to suppress
aggressive mining, the inventor should have the right to define and develop a
given field by excluding other people from the frontiers of that knowledge.
Considering the importance of industry standards in modern electronics, such a
purpose acknowledges the importance of early stage decisions that may define
the trajectory of new technological advances.

C. The Thin Protection on Software Under Copyright Law

The
Copyright Act defines a “computer program” as “a set of statements or
instructions to be used directly or indirectly in a computer to bring about a
certain result.”[19] Though it may be counterintuitive
to grant copyright protection for “useful arts” covered by patents, Congress
has explicitly mandated copyright protection for software.[20] However, as will be discussed
below, copyright protection of software has been significantly limited due to case
law.

Copyright
protects against literal infringement of the text of the program. Source code,
code lines that the programmers “author” via computer languages such as C++ and
Python, is protected under copyright as literary work.[21] In Apple v. Franklin Corp., the Third Circuit Court of Appeals held
that object code, which is the product of compiling the source code, is also
considered a literary work.[22] Given that compiled code is a
“translation” of the source code, this ruling seems to be an obvious extension
of copyright protection. Removing the copyright distinction between source code
and object code better reflects the nature of computer languages such as Perl,
where the source code is not translated into object code but rather is directly
fed into the computer for execution. However, the scope of protection on either
type of code is very narrow. The copyright system protects the author against
literal copying of code lines. This leaves open the opportunity for competitors
to avoid infringement by implementing the same algorithm using different text.

Fortunately,
in addition to protection against literal copying of code, copyright law may
provide some protection of the structure and logical flow of a program.
Equivalent to protecting the “plot” of a novel, the Second Circuit Court of
Appeals ruled that certain elements of programming structure are considered an
expression (copyrightable) rather than idea (not copyrightable), extending
copyright protection to non-literal copying.[23] The Computer Associates International v. Altai court applied a
three-step test to determine whether a computer program infringes other
programs—(1) map levels of abstraction of the program; (2) filter out
protectable expression from non-protectable ideas; and (3) compare which parts
of the protected expression are also in the infringing program.[24]

The
merger doctrine is applied to step two of the Altai test to limit what may be protected under copyright law.
Under the merger doctrine, code implemented for efficiency reasons is considered
as merged with the underlying idea, hence not copyrightable.[25] Since most algorithms are
developed and implemented for efficiency concerns, the Altai framework may prevent significant aspects of software
algorithms from receiving copyright protection. This means that for algorithms
related to computational efficiency, patents may provide significantly more
meaningful protection than copyright. The Federal Circuit, in the 2016 case McRO Inc. v. Bandai Namco Games America Inc.,
ruled that patent claims with “focus on a specific means or method that
improves the relevant technology” may still be patentable.[26] Although preemption concerns
may impede patentability, exemption of patent right by preemption is narrow
compared to that of copyright by the merger doctrine.

Scène à
faire doctrine establishes yet another limitation on copyright for computer
programs. Aspects of the programs that have been dictated by external concerns
such as memory limits, industry standards and other requirements are deemed as
non-protectable elements.[27] For mobile application
software, it is difficult to imagine programs that are not restricted by form
factors such as mobile AP computation power, battery concerns, screen size, and
RAM limitations. As for machine learning software, the algorithms determine the
“worthiness” of computation paths based on conserving computational resources.
The external factors that define the very nature and purpose of such machine
learning algorithms may exempt them from copyright protection.

D. Comparing Trade Secret and Non-disclosures with Patents

The
crucial distinction between trade secret and patent law is secrecy. While
patent applicants are required to disclose
novel ideas to the public in exchange for a government granted monopoly, trade
secret requires owners to keep the information secret. Though trade secret protection prevents outsiders from
acquiring the information by improper means, it does not protect the trade
secret against independent development or even reverse engineering of the
protected information. In trade secret doctrine, the existence of prior
disclosed art is only relevant for discerning whether the know-how is generally
known, a different and simpler analysis than the issue of novelty in patent
law.[28]

The
United States Supreme Court has specified in Kewanee Oil that all matters may be protected under trade secret
law, regardless whether it may or may not be patented.[29] The Kewanee Oil court predicted that inventors would not resort to
trade secret when offered a presumptively stronger protection by patent law:

“The
possibility that an inventor who believes his invention meets the standards of
patentability will sit back, rely on trade secret law, and after one year of
use forfeit any right to patent protection, 35 U.S.C. § 102(b), is remote
indeed.”[30]

Trade
secret is an adequate form of protection for innovators that are concerned with
the limits of what may be patentable. The secrecy requirement of trade secret
inherently provides protection that may potentially outlive any patent rights,
provided a third party does not independently acquire the secret. This
coincides with an interesting aspect of machine learning and big data—the need
for massive amounts of data. Developers need data to “train” the algorithm, and
increase the accuracy of the machine learning models. Companies that have
already acquired massive amounts of data may opt to keep their data secret,
treating the aggregated data as a trade secret.

In
addition to the amount of amassed data, companies have all the more reason to
keep their data secret if they have access to meaningful, normalized data. Even
if a company amasses an enormous amount of data, the data sets may not be
compatible with each other. Data gathered from one source may have different
reference points or methodologies that are not immediately compatible with data
from another source. This raises the concern of “cleaning” massive amounts of
data.[31] Such concerns of data
compatibility mean that parties with access to a single, homogenous source of high
quality data enjoy a significant advantage over parties that need to pull data
from multiple sources.

However,
data secrecy may not be a suitable strategy for companies that are aiming for
cross-industry data aggregation. Institutions such as Global Alliance for
Genomics and Health are promoting data sharing between research participants.
The Chinese e-commerce giant Alibaba announced a data sharing alliance with
companies such as Louis Vuitton and Samsung to fight off counterfeit goods.[32] To facilitate the development
of technology and to mitigate risks, various companies and research
institutions across diverse fields are engaging in joint development efforts
and alliances. Seeking protection under trade secret runs against this trend of
engaging in effective cross-industry collaboration. Yet there are
countervailing arguments that trade secret promotes disclosure by providing
legal remedies that can replace the protection of secrets.[33] Parties can sidestep the
limitations of trade secrets by sharing proprietary information under the
protection of contract law. While data sharing practices may void trade secret
protection, the nature of continued accumulation of data and carefully drafted
contractual provisions may provide sufficient protection for the data
owners.

II. Placing Machine Learning within Intellectual Property Law

“Learning is any process by which a system
improves performance from experience.”

–
Herbert Simon, Nobel Prize in Economics 1978.

The
concept of machine learning relates to computer programs that have the
capability to improve performance based on experience, with limited
intervention of the programmer.[34] Machine learning models have the capability to
automatically adapt and customize for individual users, discover new patterns
and correlations from large databases, and automate tasks that require some
intelligence by mimicking human intuition.[35] This section dissects the
mechanics of machine learning to identify the aspects of machine learning
innovations that are at issue as intellectual property.

A. Machine Learning Basics

Machine
learning methods are divided into two different approaches—supervised machine
learning and unsupervised machine learning. For supervised machine learning,
models are typically established by applying “labeled” sets of data to a
learning algorithm. Labeled data refers to data sets that have both relevant
features and the target results that the programmer is interested in. For
example, we may be interested in developing a machine learning model that
classifies images with dogs in them. The data sets for supervised machine
learning would indicate whether a given images has dogs or not. The learning
process begins with the algorithm fitting trends found in the training data set
into different types of models. The algorithm compares the prediction errors of
the models by inputting the validation set data into each model, measuring
their accuracy. This allows the algorithm to decide which of the various models
is best suited as the resulting machine learning model. Finally, the machine
learning model is then evaluated by assessing the accuracy of the predictive
power of the model. The developed model is then applied to data without a
correct answer to test the validity of the model. In unsupervised machine
learning, the data sets are “unlabeled” data, which may not contain the result
that the programmer is interested in. Returning to our dog image classification
example, data sets for unsupervised machine learning will have pictures of
various animals that are not labeled—the computer does not know which pictures
are associated with dogs. The unsupervised machine learning algorithm develops
a model that extracts common elements from the picture, teaching itself the set
of features that makes the subject of the picture a dog. In essence,
unsupervised machine learning uses data sets that do not have specific labels
fed into the algorithm for the purpose of identifying common trends embedded in
that data set.

The
objective of developing such machine learning models varies. Sometimes the goal
is to develop a prediction model that can forecast a variable from a data set.
Classification, which assigns records to a predefined group, is also a key
application of the algorithm. Clustering refers to splitting records into
distinct groups based on the similarity within such group. Association learning
identifies the relationship between features.

Figure 1. Overview of Machine
Learning Model Development

Figure
1 illustrates the overall process of machine learning model development. The
learning process of machine learning algorithms begins with aggregation of
data. The data originates from an array of diverse sources ranging from user
input, sensor measurement, or monitoring of user behavior.[36] The data sets are then
preprocessed. The quality of data presents a challenge in improving machine
learning models—any data that has been manually entered contains the
possibility of error and bias.[37] Even if the data is collected
through automatic means, such as health monitoring systems or direct tracking
of user actions, the data sets require preprocessing to account for systematic
errors associated with the recording device or method.[38] This includes data skews due to
difference between individual sensors, errors in the recording or transmission
of data, and incorrect metadata about the sensor.[39] Simply put, the data sets may
have differing reference points, embedded biases, or differing formats. The
“cleaning” process accommodates for the data skews.

The
objective of machine learning models is to identify and quantify “features”
from a given data set. The term “feature” refers to individually measurable
property of an observed variable.[40] From the outset, there may be an extensive list of features
that are present in a set of data. It would be computationally expensive to
define and quantify each feature, and then to identify the inter-feature
relationships, from massive amounts of data. Due to the high demand for the computational
power required for processing massive amounts of data, dedication of
computational resources to features that are outside the scope of the
designer’s interest would be a waste of such limited computational capacity.[41] The machine learning algorithm
reduces waste of computational resources by applying dimensionality reduction
to the pre-processed data sets.[42] The algorithm can identify an
optimal subset of features by reducing the dimension and the noise of the data
sets.[43] Dimensionality reduction allows
the machine learning model to achieve higher level of predictive accuracy,
increased speed of learning, and improves the simplicity and comprehensibility
of the results.[44] However, the reduction process
has limitations—reducing dimensionality inevitably imposes a limit on the
amount of insights and information that may be extracted from the data sets. If
the machine learning algorithm discerns a certain feature, the model would not
be able to draw inferences related to said feature.

Following
dimensionality reduction, the machine learning algorithm attempts to fit the
data sets into preset models. Typically, three different types of data are fed
into the machine learning model—training set, validation set, and test set.[45] The machine learning algorithm
“trains” the model by fitting the training set data into various models to
evaluate the accuracy of each selection. Then the validation set is used to
estimate error rates of each model when applied to data outside the training
set that was used to develop each model. Through this process, the machine
learning algorithm selects the model that best describes the characteristics
and trends of the target features from the test and validation sets.[46]
The test set is then used to calculate the generalized prediction error,
which is reported to the end user for proper assessment of the predictive power
of the model.[47] Simply put, the training test
and validation set is used to develop and select a model that reflects the
trends of the given data set, and the test set is used to generate a report on
the accuracy of the selected model.

The
crucial elements in developing a machine learning model are (1) training data,
(2) inventions related to the machine learning algorithm such as the method of
preprocessing the training data, the method of dimensional reduction, feature
extraction, and the method of model learning/testing, and (3) the machine
learning model and output data.[48] An ancillary element associated
with the three elements above is the human talent that is required to implement
such innovation.[49] Innovators in the field of machine learning may protect
their investments by protecting one or more of the elements listed above.

The
difference between training data and output data, as well as the difference
between the machine learning algorithm and the machine learning model, are best
illustrated with an example. Let us assume a credit card company wants to use
machine learning to determine whether the company should grant a premium credit
card to a customer. Let us further assume that the company would prefer to
grant this card to customers that would be profitable to the company while
filtering out applicants that are likely to file for bankruptcy. Data sets
about prior applicant information would correspond to training data. The company would apply a mathematical method of
extracting insight about the correlation between features and the criteria that
the company wants to evaluate (e.g., profitable for the firm or likely to file
bankruptcy). The mathematical methods are referred as machine learning algorithms. The resulting mechanism, such as a
scoring system, that determines the eligibility of card membership is the machine learning model. The credit card
applicant’s personal data would be the input
data for the machine learning model, and the output data would include information such as expected
profitability of this applicant and likelihood of bankruptcy for this
applicant.

B. Industry Trends in Machine Learning

Discussing
incentive structures and trends behind the machine learning industry is
essential in identifying adequate methods of intellectual property rights. The
current trends in the world of machine learning will predict what intellectual
property regime is most useful to companies to protect their work.

The United
States has chronically struggled to maintain adequate supply of talent in the
high-tech industry, a deficit of talent that continues in the field of machine
learning.[50] From a report by the McKinsey
Global Institute, the United States’ demand for talent in deep learning “could
be 50 to 60 percent greater than its projected supply by 2018.”[51] Coupled with the dearth of
machine learning specialists, the short employment tenure of software companies
further complicates the search for talent. Software engineers from companies
such as Amazon and Google have reported an average employment tenure of one
year.[52] While some parts of the high
attrition rate may be attributed to cultural aspects of the so-called “Gen Y”
employees, the “hot” demand for programming talent has significant impact on
the short employee tenure.[53] Job mobility within the
software industry is likely to increase as the “talent war” for data scientist
intensifies. Employee mobility and California’s prohibition against “covenants
not to compete” have been accredited as a key factor behind the success of
Silicon Valley.[54] Another trend in the field is
the rapid advances in machine learning methods. Due to the fast-paced
development of the field, data scientists and practitioners have every reason
to work with companies that would allow them to work at the cutting edge of
machine learning, using the best data sets. This may influence the attrition
rates and recruiting practices of the software industry mentioned above.[55] Eagerness of employees to
publish scientific articles and contribute to the general machine learning
committee may be another factor of concern.

To
accelerate innovation by repurposing big data for uses different from the
original purpose, and to form common standards for machine learning, more
industries are joining alliances and collaborations.[56] Cross-industry collaborations
may enable endless possibilities. Imagine the inferences that may be drawn by
applying machine learning methods to dietary data from home appliances,
biometric data, and data on the weather patterns around the user. Putting
privacy nightmares aside, machine learning with diverse data sets may unlock
applications that were not previously possible. More companies are attempting
to capitalize on commercial possibilities that data sharing may unlock.[57]

C. Machine Learning Innovators – Protect the Data or Inventions?

Though
it may seem intuitive that patent protection may be the best option,
innovations in machine learning may not need
patent protection. Trade secret protection on the data sets may be sufficient
to protect the interests of practicing entities while avoiding disclosure of
their inventions during the patent prosecution process. Furthermore, numerous
software patents have been challenged as unpatentable abstract subject matter
under 35 U.S.C. §101 since the Alice
decision in 2014.[58] Though subsequent decisions
provided guidelines for types of software patents that would survive the Alice decision, it is not clear how the
judiciary will view future machine learning patents. Such issues raise the
question about the patentability of machine learning – should we, and can we,
resort to patents to protect machine learning inventions?

Following
the discussion on the building blocks of machine learning and recent emerging
trends in the field, this section discusses the mode and scope of protection
that current legal system provides for each element pertinent to innovation in
machine learning. The possible options for protecting innovations are (1)
non-disclosure agreements and trade secret law, (2) patent law, and (3)
copyright. The three options for protection may be applied to the three primary
areas of innovation—(1) training data, (2) inventions related to computation,
data processing, and machine learning algorithms, and (3) machine learning
models and output data. This discussion will provide context about the methods
of protection for innovations in machine learning by examining the costs and benefits
of the various approaches.

1. Protecting the Training Data—Secrecy Works Best

Access
to massive amounts of training data is a prime asset for companies in the realm
of machine learning. The big data phenomenon, which triggered the surge of
interest in machine learning, is predicated on the need for practices to analyze
large data resources and the potential advantages from such analysis.[59] Lack of access to a critical
mass of training data prevents innovators from making effective use of machine
learning algorithms.

Previous
studies suggest that companies resent sharing data with each other.[60] Michael Mattioli discusses the hurdles against sharing data
and considerations involved with reuse of data in his article Disclosing Big Data.[61] Indeed, there may be practical issues that prevent recipients of data from engaging in data
sharing. Technical challenges in comparing data from different sources, or
inherent biases embedded in data sets may be reasons that complicate receiving
outside data.[62] Mattioli also questions the
adequacy of the current patent and copyright system to promote data sharing and
data reuse—information providers may
prefer not to disclose any parts of their data due to the rather thin legal
protection for databases.[63]

Perhaps
this is why secrecy seems to be the primary method of protecting data.[64] The difficulty of reverse
engineering to uncover the underlying data sets promotes the reliance on
non-disclosure.[65] Compared to the affirmative
steps required to maintain trade secret protection if the data is disclosed,
complete non-disclosure may be a cost effective method of protecting data.[66] Companies that must share data
with external entities may exhibit higher reliance on contract law rather than
trade secret law. In absence of contract provisions, it would be a challenge to
prove that the trade secret has been acquired by misappropriation of the
recipient party.

The
“talent war” for data scientists may also motivate companies to keep the
training data sets secret. With a shortage of talent to implement machine
learning practices and rapid developments in the field, retaining talent is
another motivation for protecting against unrestricted access to massive
amounts of data. Companies may prefer exclusivity to the data sets that
programmers can work with — top talents in machine learning are lured to companies
with promises of exclusive opportunities to work with massive amounts of data.[67] The rapid pace of development
in this field encourages practitioners to seek opportunities that provide the
best resources to develop their skill sets. This approach is effective since a
key limitation against exploring new techniques in this field is the lack of
access to high quality big data. Overall, secrecy over training data fits well
with corporate recruiting strategies to retain the best talents in machine
learning.

Non-disclosure
and trade secret protection seems to be the best mode of protection. First,
despite the additional legal requirements necessary to qualify as trade
secrets, trade secret protection fits very well with non-disclosure strategy.
On the other hand, patent law is at odds with the principle of non-disclosure.
While trade secret law provides companies protection without disclosing
information, patent law requires disclosure in exchange for monopolistic
rights. Furthermore, neither patent nor copyright provide adequate protection
for underlying data. Patent law rewards creative concepts and inventions, not
compiled facts themselves. Copyright may protect labeling or distinct ways of
compiling information, but does not protect underlying facts. Also, as a
practical matter, the difficulty of reverse engineering of machine learning
models does not lend well to detecting infringement. Analysis of whether two
parties used identical training data would not only be time consuming and
costly, but may be fundamentally impossible.

If
companies were to seek protection of training data, it would be best to opt for
secrecy by non-disclosure. This would mean companies would opt out of the
cross-industry collaborations that were illustrated above. This may be less of
a concern for innovation, as companies may still exchange output data as means
of facilitating cross-industry collaboration.

2. Protecting the Inventions—Patent Rights Prevail

Adequate
protection over inventive approaches in processing data is becoming
increasingly important as various industries begin to adopt a collaborative
alliance approach in machine learning. Cross-industry collaboration requires
implementation of methods such as preprocessing diverse data sets for
compatibility. As the sheer amount of data increases, more processing power is
required. The machine learning algorithm needs to maintain a high degree of
dimensionality to accurately identify the correlations between a high number of
relevant features. The need for more innovative ideas to address such
technological roadblocks will only intensify as we seek more complex
applications for machine learning.

The
three primary areas where novel ideas would facilitate innovations in machine
learning are pre-training data processing, dimensional reduction, and the
machine learning algorithm.

Access
to massive amounts of data alone is not sufficient to sustain innovation in
machine learning. The raw data sets may not be compatible with each other,
requiring additional “cleaning” of data prior to machine learning training.[68] The data provided to the
machine learning algorithm dictates the result of the machine learning model,
hence innovations in methods to merge data with diverse formats is essential to
enhancing the accuracy of the models. As cross-industry data analysis becomes
more prominent, methods of merging data will have more significant impact on
advancing the field of machine learning than mere collection of large data
sets. Cross-industry data sharing would be useless unless such data sets are
merged in a comparable manner.[69]

Companies can opt to protect their inventive methods by
resorting to trade secret law. The difficulty of reverse engineering machine
learning inventions, coupled with the difficulty of patenting software methods
provides incentives for innovators to keep such inventions secret from the
public. However, two factors would render reliance on non-disclosure and trade
secret ineffective—frequent turnover of software engineers and rapid speed of
development in the field.

Rapid
dissemination of information from employment mobility may endanger intellectual
property protection based on secrecy. Furthermore, while the law will not
protect former employees that reveal trade secrets to their new employers, the
aforementioned fluid job market coupled with general dissemination of
information make it difficult to distinguish between trade secrets from former
employment and general knowledge learned through practice. The difficulties of
reverse engineering machine learning models work against the trade secret owner
as well in identifying trade secret misappropriation—how do you know others are
using your secret invention? The desire for software communities to discuss and
share recent developments in the field does not align well with the use of
secrecy against innovations in machine learning. Secrecy practices
disincentivize young data scientists from joining due to the limits against
gaining recognition.[70]

The
rapid development of machine learning technology also presents challenges
against reliance on trade secret law.
Secret methods may be independently developed by other parties. Neither
trade secret law nor non-disclosure agreements protect against independent
development of the same underlying invention.[71] Unlike training data, machine
learning models, or the output data, there are no practical limitations that
impedes competitors from independently inventing new computational methods of
machine learning algorithms.

With
such a fluid employment market, high degree of dissemination of expertise, and
rapid pace of development, patent protection may provide the assurance of
intellectual property protection for companies developing inventive methods in
machine learning. Discussions on overcoming the barriers of patenting software
will be presented in later sections.[72]

3. Protecting the Machine Learning Models and Results—Secrecy Again

The two
primary products from applying the machine learning algorithms to the training
data are the machine learning model and the accumulation of results produced by
inputting data into the machine learning model. The “input data” in this
context may refer to individual data that is analyzed by the insights gained
from the machine learning model.

In a
recent article, Brenda Simon and Ted Sichelman discuss the concerns of granting
patent protection for “data-generating patents,” which refers to inventions
that generate valuable information in their operation or use.[73] Exclusivity based on patent protection may be extended
further by trade secret protection over the data that has been generated by the
patented invention.[74] Simon and Sichelman argue that
the extended monopoly over data may potentially overcompensate inventors since
the “additional protection was not contemplated by the patent system[.]”[75] Such expansive rights will
cause excessive negative impact on downstream innovation and impose exorbitant
deadweight losses.[76] The added protection over the
resulting data derails the policy rationale behind the quid pro quo exchange
between the patent holder and the public by excluding the patented information
from public domain beyond the patent expiration date.[77]

The
concerns addressed in data-generating patents also apply to machine learning
models and output data. Corporations may obtain patent protection over the
machine learning models. Akin to a preference for secrecy for training data,
non-disclosure would be the preferred mode of protection for the output data.
The combined effect of the two may lead to data network effects where users
have strong incentives to continue the use of a given service.[78] The companies that have
exclusive rights over the machine learning model and output data gather more
training data, increasing the accuracy of their machine learning products. The
reinforcement by monopoly over the means of generating data allows few companies
to have disproportionately strong dominance over their competitors.[79]

Market
dominance by data-generating patents becomes particularly disturbing when the
patent on a machine learning model preempts other methods in the application of
interest. Trade secret law does not provide protection against independent
development. However, if there is only one specific method to obtain the best
output data, no other party would be able to create the output data
independently. The exclusive rights over the only methods of producing data
provides means for the patent holder to monopolize both the patent and the
output data.[80] From a policy perspective, the
excessive protection does seem troubling. Yet such draconian combinations are
less feasible after the recent rulings on patentable subject matter of
software, which will be discussed below.[81] Mathematical equations or
concepts are likely directed to an “abstract concept,” thus will be deemed
directed to a patent ineligible subject matter.[82] Furthermore, though recent
cases in the Federal Circuit have found precedents where software patents
passed the patentable subject matter requirement, those cases expressed
limitations against granting patents that would improperly preempt all
solutions to a particular problem.[83] The rapid pace of innovation in
the field of machine learning compared to the rather lengthy period required to
obtain patents may also dissuade companies from seeking patents. Overall,
companies have compelling incentives to rely on non-disclosure and trade
secrets to protect their machine learning models instead of seeking patents.

The
secrecy concerns regarding training data applies to machine learning models and
the output data as well. Non-disclosure would be the preferred route of
obtaining protection over the two categories. However, use of non-disclosure or
trade secrets to protect machine learning models and output data presents
challenges that are not present in the protection of training data. The use of
secrecy to protect machine learning models or output data conflicts with
recruiting strategies to hire and retain top talent in the machine learning
field. The non-disclosure agreements limit the employee’s opportunity to gain
recognition in the greater machine learning community. In a rapidly developing
field where companies are having difficulty hiring talent, potential employees
would not look fondly on corporate practices that limit avenues of building a
reputation within the industry.[84]

Companies have additional incentives to employ a rather lenient
secrecy policy for machine learning models and the output data. They have
incentives to try to build coalitions with other companies to monetize on the
results. Such cross-industry collaboration may be additional source of income
for those companies. The data and know-how that Twitter has about fraudulent
accounts within their network may aid financial institutions such as Chase with
novel means of preventing wire fraud. The reuse of insights harvested from the
large amount of raw training data can become a core product the companies would
want to commercialize. Data reuse may have an incredible impact even for
applications ancillary to the primary business of the company.

Interesting
aspects of disclosing machine learning models and output data are the
difficulty of reverse engineering and consistent updates. If the company
already has sufficient protection over the training data and/or the
computational innovations, competitors will not be able to reverse engineer the
machine learning model from the output data. Even with the machine learning
model, competitors will not be able to provide updates or refinements to the
model without the computational techniques and the sufficient data for training
the machine learning algorithm. In certain cases, the result data becomes
training data for different applications, which raises concerns of competitors
using the result data to compete with the innovator. Yet the output data would
contain less features and insights compared to the raw training data that the innovator
possesses, and therefore would inherently be at a disadvantage when competing
in fields that the innovator has already amassed sufficient training data.

Grant
of patents on machine learning models may incentivize companies to build an
excessive data network while preempting competitors from entering competition.
This may not be feasible in the future, as technological preemption is becoming
a factor of consideration in the patentable subject matter doctrine. Companies
may use secrecy as an alternative, yet may have less incentives to keep secrecy
compared to the protection of training data.

D. Need of Patent Rights for Machine Learning Inventions in the Era of Big Data

The
current system, on its surface, does not provide adequate encouragement for
data sharing. If anything, companies have strong incentives to avoid disclosure
of their training data, machine learning model, and output data.

Despite
these concerns, data reuse may enable social impacts and advances that would
not be otherwise possible. Previous studies have pointed out that one of the
major barriers preventing advances in machine learning is the lack of data
sharing between institutions and industries.[85] Data scientists have
demonstrated that they were able to predict flu trends with data extracted from
Twitter.[86] Foursquare’s location database
provides Uber with the requisite data to pinpoint the location of users based
on venue names instead of addresses.[87] Information about fraudulent
Twitter accounts may enable early detection of financial frauds.[88] The possibilities that
cross-industry data sharing may bring are endless.

To
encourage free sharing of data, companies should have a reliable method of
protecting their investments in machine learning. At the same time, protection
based on non-disclosure of data would defeat of purpose of promoting data
sharing. Hence protection over computation methods involved with machine
learning maintains the delicate balance between promoting data sharing and
protecting innovation.

Protection
over inventions in the machine learning algorithm provides one additional merit
other than allowing data sharing and avoiding the sort of excessive protection
that leads to a competitor-free road and data network effects. It incentivizes
innovators to focus on the core technological blocks to the advancement of
technology, and encourages disclosure of such know-how to the machine learning
community.

Then what are the key obstacles in obtaining patents in
machine learning inventions? While there are arguments that the definiteness
requirement of patent law is the primary hurdle against patent protection of
machine learning models due to reliance on subjective judgment, there is no
evidence that the underlying inventions
driving big data faces the same challenge.[89] Definiteness may be overcome by
providing reasonable certainty for those skilled in the art of defining what
the scope of the invention is at the time of filing.[90] There is no inherent reason why
specific solutions for data cleaning, enhancement of computation efficiency,
and similar inventions would be deemed indefinite by nature.

Since
the United States Supreme Court invalidated a patent on computer implemented financial
transaction methods in the 2014 Alice
decision, the validity of numerous software and business method patents were
challenged under 35 U.S.C. §101.[91] As of June 8^th,
2016, federal district courts invalidated 163 of the 247 patents that were
considered under patentable subject matter—striking down 66% of challenged
patents.[92] The U.S. Court of Appeals for
the Federal Circuit invalidated 38 of the 40 cases it heard.[93]

Arguably,
the public benefits more from such high rates of post-issuance invalidity. The
public still has access to the disclosures from the patents and patent
applications. In reliance on granted patents, companies may have already
invested in growing related businesses, catering to the need of consumers. At
the same time, the patent holder’s monopolistic rights have been shortened as
the result of litigation. Effectively, the price that the public pays to
inventors in exchange for the benefits of disclosure is reduced.

Yet the
high degree of invalidity raises several concerns for the software industry.
Smaller entities, lacking market influence and capital, have difficulty
competing against established corporations without the monopolistic rights
granted through the patent system. Investors become hesitant to infuse capital
into startups for fear that invalidity decreases the worth of patents. Reliance
on trade secret has its own limitations due to the disclosure dilemma—the
inventor needs to disclose the secret to lure inventors, but risks losing
secrecy in the process. Copyright law does not provide appropriate protection.
The restrictions imposed by the merger doctrine and scène à
faire doctrine
constrain copyright protection of software. Though copyright provides an
alternative method of protecting literal copying of code, it does little to
protect the underlying software algorithms and innovation.

Ultimately,
the increase of alliances and collaboration provides incentives for parties to
obtain patent rights. Reliance on trade secret or copyright are not suitable
methods of protecting their intellectual property. Furthermore, market power or
network effects alone cannot sufficiently mitigate the risks involved with
operating a business. Patents become even more important for startups since
patents provide investors with assurance that in the worst case, the patents
may still serve as potential collateral.

III. Patentability of Machine Learning Innovations in the Era of Big Data

Patentable subject matter continues to be a barrier for
patenting innovations in software. Additional doctrines such as enablement,
written description, and obviousness are also serious obstacles against
obtaining patents, yet such requirements are specific to each claimed invention
and the draftsmanship of claims. Subject matter is considered a broader, categorical
exclusion of patent rights. This section explores the current landscape of the
patentable subject matter doctrine in the software context.

A. Alice: The Legal Framework of Patentable Subject Matter in Software

The
complexity involved with software, coupled with the relatively broad scope of
software patents, has presented challenges in identifying the boundaries of the
claims.[94] Many members of the software community detest imposing
restrictions on open source material and attest that many key innovations in
algorithms are rather abstract.[95] Such hostility against
patenting software has raised the question of whether patent rights should be
the proper method of protecting innovations in software.

Alice was a case that embodied such
opposition to the grant of software patents. The case involved patents on
computerized methods for financial trading systems that reduce “settlement
risk” when only one party to financial exchange agreement satisfies its
obligation.[96] The method proposed the use of
a computer system as a third-party intermediary to facilitate the financial
obligations between parties.[97] The United States Supreme Court
ruled that the two-step test established from Mayo governed all patentable subject matter questions.[98] In particular, for the abstract
idea context, the Supreme Court established the following two-step framework
for patentable subject matter of software inventions:

1. Step
one: “[D]etermine whether the claims at issue are directed to a patent-ineligible concept. If so, the Court then asks
whether the claim’s [additional] elements, considered both individually and ‘as
an ordered combination,’ ‘transform the nature of the claim’ into a
patent-eligible application.”[99]

2. Step
two: “[E]xamine the elements of the claim to determine whether it contains an ‘inventive concept’ sufficient to
‘transform’ the claimed abstract idea into a patent-eligible application. A
claim that recites an abstract idea must include ‘additional features’ to
ensure that the [claim] is more than a drafting effort designed to monopolized
the [abstract idea]” which requires “more than simply stat[ing] the [abstract
idea] while adding the words ‘apply it.’”[100]

The Alice Court found that the patent on
financial transaction was “directed to a patent-ineligible concept: the
abstract idea of intermediated settlement,” and therefore failed step one.[101] Furthermore, the Court ruled
that the claims did “no more than simply instruct the practitioner to implement
the abstract idea of intermediated settlement on a generic computer” and did
not provide an inventive concept that was sufficient to pass step two.[102]

B. The post-Alice cases from the Federal Circuit

The Alice framework was considered as a huge
setback for the application of patentable subject matter doctrine to software.
It was a broad, categorical exclusion of certain inventions that were deemed
“directed to” an abstract idea, natural phenomenon, or law of nature. The
biggest misfortune was the lack of guidance in the Alice decision on the threshold for such categorical exclusion—we
were left without any suggestions on the type of software patents that would be
deemed as patentable subject matter.

The
recent line of cases in the Federal Circuit provides the software industry with
the much-needed clarification on the standards that govern patentability of
software inventions. Enfish v. Microsoft,
decided on March 2016, involved a “model of data for a computer database
explaining how the various elements of information are related to one another”
for computer databases.[103] In June 2016, the Federal
Circuit decided another case on the abstract idea category for patentable
subject matter. Bascom Global v. AT&T
Mobility is on a patent disclosing an internet content filtering system located
on a remote internet service provider (ISP) server.[104] Shortly after Bascom, the Federal Circuit decided McRO v. Bandai Namco Games in September
2016.[105] The case ruled that an
automated 3D animation algorithm that renders graphics in between two target facial
expressions is patentable subject matter.[106]

The rulings from the Federal Circuit on the aforementioned
three cases provide guidelines along the two-step Alice test of patentable subject matter. The software patents in Enfish and McRO were deemed “directed to” a patent eligible subject matter,
informing the public of what may pass the first set of the Alice test. Bascom failed
the first step.[107] Yet the court ruled that those
patents had inventive concepts sufficient to transform a patent ineligible
subject matter into a patent eligible application. Combined together, the three
cases give more certainty in what may pass the 35 U.S.C. §101 patentable
subject matter inquiry.

Reiterating
the Alice test, whether an invention
is a patentable subject matter is determined by a two-step process—(1) is the
invention directed to, rather
than an application of, an abstract idea, natural phenomenon, or law of nature,
and even if so, (2) do the elements of the claim, both individually and
combined, contain an inventive concept
that transforms this invention into a patent-eligible application? The Federal
Circuit fills in the gaps that were left unexplained from the Alice ruling.

1. The Federal Circuit’s Standard for Alice Step One

The Enfish court discussed what constitutes
an abstract idea at the first step of the Alice
inquiry. Judge Hughes instructs us to look at whether the claims are directed
to a specific improvement rather than an abstract idea. In this case, the
patent provides the public with a solution to an existing problem by a
specific, non-generic improvement to computer functionality. The Enfish court ruled that such invention
is patent eligible subject matter.[108]

McRO also ruled that the facial
graphic rendering for 3D animation was not an abstract concept. Here, the
Federal Circuit again emphasized that a patent may pass step one of the Alice test if the claims of the patent
“focus on a specific means or method that improves the relevant technology.”[109] The McRO court also noted that preemption concerns may be an important
factor for the 35 U.S.C. §101 subject matter inquiry—that improper
monopolization of “the basic tools of scientific and technological work” is a
reason why such categorical carve outs against granting patents on abstract
ideas exist.[110]

Bascom provides the standards on what
would fail step one of the Alice
patentable subject matter inquiry. If the patent covers a conventional, well-known
method in the field of interest, then the invention would be considered
abstract. This is akin to the inventive concept considerations conducted at the
second phase of the 35 U.S.C. §101 subject matter inquiry.

The
main takeaway from Enfish and McRO is that in the first step of the Alice test, a patent application is not
directed to an abstract idea if (1) the invention addresses an existing problem
by specific improvements rather than by conventional, well-known methods and
(2) the claims do not raise preemption concerns. This encourages practitioners
to define the problem as broadly as possible, while defining the scope of
improvement in definite terms.

2. The Federal Circuit’s Standard for Alice Step Two, and the Overlap with Step One

The
second step of the Alice test is an
inquiry of whether the patent application, which is directed to a patent
ineligible subject, still contains a patent-worthy inventive concept. Bascom ruled in favor of granting the
patent following the second step of the Alice
test.[111] While the patent at hand was
considered directed to patent ineligible subject matter, the Bascom court found that the content
filter system invention still had an inventive concept worthy of a patent.[112] Even if elements of a claim are
separately known in prior art, an inventive concept can be found in the
non-conventional and non-generic arrangement of known, conventional pieces.
This inquiry seems like a lenient standard compared to the 35 U.S.C. §103
obviousness inquiry; hence, it is not clear if this step has an independent
utility for invalidating or rejecting a patent. Nonetheless, the court found
that merely showing that all elements of a claim were already disclosed in
prior art was not sufficient reason to make an invention patent ineligible.[113]

While
it is possible to infer sufficient reasons of ruling out inventive concepts
from the Bascom case, it is still
unclear what would warrant an invention to pass the second step of the Alice test. Cases such as DDR Holdings v. Hotels.com have
suggested that the second step of Alice
is satisfied since it involved a solution to a specific technological problem
that “is necessarily rooted in computer technology in order to overcome a
problem specifically arising in the realm of computer networks.”[114]

This
interpretation of inventive concept
becomes perplexing when comparing the two steps of Alice—both steps look to whether the proposed solution addresses
problems that are specific to a given field of interest. While we would need
additional cases to gain insight on whether the two steps have truly distinct
functions, at the very least the Federal Circuit provided essential guidelines
on what may be deemed as patentable software.

C. Applying Patentable Subject Matter to Machine Learning Inventions

As the Bascom court has taught, the first step
in the Alice inquiry is to ask
whether an invention (1) provides a solution to an existing problem by (2) a
specific, non-generic improvement that (3) does not preempt other methods of
solving the existing problem. Applying this test to inventions in machine
learning, mathematical improvements
and computational improvements would
be treated differently.

As
mentioned before, a key aspect of machine learning is the “noise” associated
with the data sets.[115] Another concern is the fitting
of a given algorithm to a certain model. Methods that facilitate the
computations of the training process may be deemed as a specific improvement.
However, machine learning algorithms themselves, including the base models that
the algorithm fits the training into would not be pertinent to just a specific
improvement. Hence, generic mathematical methods applicable to various problems
are directed to an abstract idea. For example, an invention that addresses the
issue of normalizing data from different sources would be a computational issue
and hence would pass the Alice test
given that it did not preempt other solutions to the problem of data
normalization. On the other hand, a specific mathematical equation that serves
as a starting model for the machine learning algorithm would be mathematical
and hence directed to an abstract idea. Even if the mathematical starting model
is only good for a specific application, the model is not a specific
improvement pertinent to that application. Although the model may not
necessarily be a good starting model for other applications, it is nonetheless
a generic solution that applies to other applications as well.

Conclusion

While
highly restrictive, the guidelines from the Federal Circuit still allow the
grant of patent rights for the computational aspects of machine learning
algorithms. The guidelines also would prevent highly preemptive mathematical
innovations, including data-generating patents such as machine learning models.

The
narrow range of patentability makes a patent regime appealing for computational
methods. The recent emphasis on preemption concerns acts in favor of preventing
data network effects based on data-generating patents. While not discussed in
this paper, other patentability requirements such as obviousness or
definiteness would further constraint the grant of overly broad data-generating
patents.

Such an
approach strikes the appropriate balance between promoting innovation and
encouraging data reuse for societal benefits. Compared to other approaches of
providing protection over innovations in machine learning, the narrowly
tailored approach for patent rights for computational inventions fits best with
the policy goal of promoting innovation through data reuse. The industry trends
in collaboration and recruiting also matches the proposed focus on patent law
protection.