Q&A
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
Technical Aspects of Artificial Intelligence:
An Understanding from an
Intellectual Property Law Perspective
Version 1.0, October 2019
Research Group on the Regulation of the Digital Economy
Josef Drexl, Reto M. Hilty, Francisco Beneke, Luc Desaunettes, Michèle Finck,
Jure Globocnik, Begoña Gonzalez Otero, Jörg Hoffmann, Leonard Hollander,
Daria Kim, Heiko Richter, Stefan Scheuerer, Peter R. Slowinski, Jannick Thonemann
Max Planck Institute for Innovation and Competition Research Paper Series
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
1
Technical Aspects of Artificial Intelligence:
An Understanding from an
Intellectual Property Law Perspective
Josef Drexl, Reto M. Hilty, Francisco Beneke, Luc Desaunettes, Michèle Finck,
Jure Globocnik, Begoña Gonzalez Otero, Jörg Hoffmann, Leonard Hollander,
Daria Kim, Heiko Richter, Stefan Scheuerer, Peter R. Slowinski, Jannick Thonemann
The purpose of this Q&A paper is to provide an overview of artificial intelligence* with a
special focus on machine learning* as a currently predominant subfield thereof. Machine
learning-based applications have been discussed intensely in legal scholarship, including in
the field of intellectual property law, while many technical aspects remain ambiguous and
often cause confusion.
This text was drafted by the Research Group on the Regulation of the Digital Economy of the
Max Planck Institute for Innovation and Competition in the pursuit of understanding the
fundamental characteristics of artificial intelligence, and machine learning in particular, that
could potentially have an impact on intellectual property law. As a background paper, it
provides the technological basis for the Group’s ongoing research relating thereto.
The current version summarises insights gained from background literature research,
interviews with practitioners and a workshop held in June 2019, in which following experts
in the field of artificial intelligence participated: Dr. Maximilian Alber (Charité Berlin),
Dr. Cristian Ramirez Atencia (Otto von Guericke University Magdeburg), Dr. Hoda Heidari
(ETH Zürich), Dr. Jelena Mitrović (University of Passau), Dr. Cigdem Turan (Technical
University of Darmstadt) and Heiner Zille (Otto von Guericke University Magdeburg).
Since machine learning is a dynamic field, this document reflects the current understanding
and might therefore be updated in the future. Readers are kindly invited to make further
suggestions or to provide clarifications. The presentation of the members of the Research
Group including their contact details, and the Group’s other research outputs can be found
at the following address: https://www.ip.mpg.de/link/regulation-of-the-digital-
economy.html.
Terms with an asterisk (*) are defined in the Glossary at the end of this document.
Suggested citation:
Drexl, Hilty et al., Technical Aspects of Artificial Intelligence: An Understanding from an
Intellectual Property Law Perspective, Version 1.0, October 2019, available at:
https://ssrn.com/abstract=3465577.
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
2
Contents
1. Overview of Artificial Intelligence and Machine Learning............................................. 3
Q1 What is generally understood by artificial intelligence? How does it differ from
related fields? What subcategories does it encompass?...................................................... 3
2. Machine Learning.............................................................................................................. 4
General............................................................................................................................... 4
Q2 What factors contribute to the current widespread usage of machine learning?......... 4
Q3 What are the basic characteristics of the machine learning process?.............................. 4
Training Process and Model ............................................................................................. 5
Q4 What is a machine learning model and what are its components?................................... 5
Q5 How is a model trained? What kind of human input does this necessitate? Which are
the different training methods? .................................................................................................... 7
Machine Learning Outputs ............................................................................................... 9
Q6 What are the possible outputs of a machine learning model, and to what extent are
they explainable and interpretable?............................................................................................ 9
Reverse Engineering of Machine Learning Components .............................................10
Q7 Is it possible to reverse engineer different components of the machine learning
process?...............................................................................................................................................10
Machine Learning ‘Intelligence’.....................................................................................10
Q8 Which parts of the machine learning process are pre-determined by humans?.........10
Q9 To what extent is a machine learning model a ‘black box’ that cannot be explained
by a human?.......................................................................................................................................11
3. Evolutionary Algorithms.................................................................................................11
Q10 What are evolutionary algorithms? Do they differ from machine learning? ................11
Glossary ......................................................................................................................................12
Endnotes.....................................................................................................................................14
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
3
1. Overview of Artificial Intelligence and
Machine Learning
What is generally understood by artificial intelligence?
How does it differ from related fields?
What subcategories does it encompass?
Artificial Intelligence
Artificial intelligence* is a branch of computer science.1
It is often described as computerbased
systems that are developed to mimic human behaviour.
Artificial intelligence is a catch-all term that covers machine learning*, evolutionary
algorithms* (see Q10), and other technologies like rule-based systems.2
However, due to the
fact that different subfields often overlap, the exact delineation is difficult and a subject of
controversy among researchers.
Delineation between Artificial Intelligence, Robotics, and Statistics
One of the closest neighbouring fields of artificial intelligence is robotics, which is a branch
of engineering.3
While in principle both fields are independent, overlaps emerge when
artificial intelligence is embedded in robots with the aim to enable them to better react to a
complex, uncertain and dynamic environment.4
A further neighbouring field is statistics. It is disputed to what extent machine learning
should be considered as an application of statistics. Indeed the purposes of both fields might
not be identical: while statistics aims at analysing distributions or correlations, artificial
intelligence is more about behaving intelligently or predicting (the future) correctly. Still,
machine learning process relies heavily on statistical methods.
Machine Learning and Evolutionary Algorithms as Subfields of Artificial
Intelligence
Machine learning is currently the most commonly used subfield of artificial intelligence. It
deals with teaching a computer program to identify patterns in data and to apply the
knowledge to new data.5
One of the most advanced subfields of machine
learning is deep learning* (see Q4).
While this document focuses mainly on machine
learning, it is important to also consider other
types of artificial intelligence that are gaining
importance, such as evolutionary algorithms
(see Q10).
Machine learning
Deep
learning
Evolutionary
algorithms
Artificial
intelligence
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
4
2. Machine Learning
General
What factors contribute to the current widespread usage
of machine learning?
Machine learning* can be described as a general-purpose technique as it can be applied in
all sectors to optimise decision-making and facilitate innovation.
Breakthroughs in mathematical optimisation methods occurred already in the 1980s. Since
then, there have been further advancements in how these methods are applied in machine
learning (e.g. generative adversarial networks*, deep neural networks*).6
The current
widespread penetration of machine learning is mostly attributed to two major factors: the
emergence of large data sets that can be used for the training process, and the increase in
computing power.7
What are the basic characteristics of the machine
learning process?
The entire process relies on the analysis of data by different algorithms*. An algorithm is a
step-by-step instruction. Generally, algorithms are encoded as software* to enable their
readability by computers.
Machine learning* processes exist in different variations, depending on the data they build
upon, and their task. In order to make further explanations more understandable, a basic
machine learning process relying on supervised learning* (see Q5) and an artificial neural
network* as model architecture* (see Q4) is presented below. Unless stated differently, this
document uses this type of machine learning as its basis.
Machine learning consists of several stages. First, a model architecture is programmed;
second, a model* is developed through the training process based on a training algorithm*
and training data sets*; third, the model is applied to new data to generate a certain output
(e.g. a prediction, see Q6).
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
5
Example: A machine learning model might be used to recognise cats in pictures. In the case of
supervised machine learning, the model is trained on a data set containing labelled data (i.e., each
picture is accompanied by the information whether there is a cat in the picture), allowing it to
become more accurate. Once the training is completed, the model should in principle be capable
of recognising from an unlabelled picture whether a cat appears in it (output). This model could
finally be implemented in a self-driving car, allowing it, for instance, to brake when confronted
with a cat (application).
Training Process and Model
What is a machine learning model and what are its
components?
Model
The trained machine learning* model* is the immediate output of the training process. It is
an algorithm* based upon a (nonlinear) mathematical function that generates output based
on the learned patterns in the training data*.
One type of models are artificial neural networks*, the structure of which imitates the
functioning of a human brain. These models rely on an architecture* which is usually
established by a programmer prior to the training process and is composed of layers of
neurons* connected by weights*. Each neuron is a mathematical function which transforms
inputs (the numeric value of the upstream weights) into an output (the numeric value of the
downstream weights). The model is composed of the sum of all the functions entailed in the
neurons.
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
6
When the number of layers is high, the neural network* is described as deep (deep neural
network), whereby the scientific community does not agree upon a clear delineation.
Trainable Parameters and Hyperparameters
Two types of parameters might be distinguished in the machine learning context, namely
hyperparameters* and trainable parameters*. Trainable parameters evolve during the
training process, whereas hyperparameters are fixed before the training process and do not
evolve.
The weights* are trainable parameters. They are numeric values that are first randomly
allocated and then optimised during the training process.
Alternatively, a model trained for a different, yet similar task can be used as a starting point
instead of randomly allocating the weights (transfer learning*). This allows a quicker
optimisation compared to starting from scratch, but a training process remains unavoidable.
The architecture*, on the contrary, is a hyperparameter, which means that it (currently) needs
to be developed by a human before the training process and does not evolve during it. The
determination of the optimal structure relies on heuristic* methods and human know-how.
Research is nonetheless being conducted to allow the modification of the architecture
through the training process.
Online and Offline Models
A model can be online or offline. In the case of the currently predominant offline models*
(also called static models), the optimisation process and the actual application are separated.
In the application phase, no modification of the weights takes place.
If the model is designed as an online model* (also called dynamic model), its optimisation
never ends, because even when being applied, its output is used to continually modify the
weights. This requires feedback on the correctness of the output.
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
7
How is a model trained? What kind of human input does
this necessitate? Which are the different training
methods?
Training Process
In the training process, training data* is ‘fed’ into the model*. Based on this, the training
algorithm* (sometimes also called optimisation algorithm) optimises trainable parameters*.
A training algorithm includes a loss function* that reflects the accuracy of the model. A loss
function is a mathematical function which evaluates the magnitude of error of a model. The
smaller the loss function, the better the model is. The training algorithm will, therefore, seek
to minimise the loss function by finding the best combination of trainable parameters.
The modification of the weights* is hence a mathematical operation. Therefore, under
identical parameters, an identical model would be developed. Slight changes of the
parameters (like for instance regarding the order of feeding the training data, or in the initial
allocation of the weights, see Q4) would lead to the creation of a model that is slightly
different but likely capable of producing outputs of similar accuracy.
Development, Availability, and Choice of Training Algorithms
Training algorithms* vary in originality from standard to uniquely developed. Many
established training algorithms are standard, and available online in open-source libraries in
the form of pre-written software*. However, for some problems, new algorithms* have to be
developed, necessitating human and financial investments.
The choice of particular training methods requires know-how and currently relies on certain
heuristic methods*. Heuristics is an approach to problem-solving relying on experience and
intuition rather than a pure scientific methodology. Heuristic methods are often used due to
the lack of sufficient computing power or the absence of exact methods for the solving of
certain problems. If computational power was limitless, it would be possible to test every
possible solution to find out which is the best one. Research is being conducted to replace
heuristic methods by exact ones. However, even though available computing power is
continuously increasing, experts doubt that in near future all heuristic methods could be
replaced by exact ones.
Supervised, Unsupervised, and Reinforcement Learning
Three training methods can be differentiated regarding the type and the way training data
is used.
In the case of supervised learning*, the model is ‘told’ during the optimisation process what
the training data it is confronted with represents. Hence, it knows whether the prediction it
made in the optimisation process was right or wrong. For doing that, it necessitates labelled
data, for which additional investments (for instance human involvement) are required.
Supervised learning is currently the most common form of machine learning*.
Example: Cat recognition (see above, Q3).
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
8
Unsupervised learning*, on the contrary, relies on unlabelled data. Here, a model is trained
to identify similarities, parallels and/or differences in data, the main use case being
clustering. Since labelling of the training data is not required, this type of training requires
less human participation. On the other hand, a more extensive human interpretation of the
output is required.
Example: An algorithm* analyses data about customers to build groups based on their potential
purchasing power to customise the offers.
Reinforcement learning* does not rely on pre-existing data sets, but rather gathers data from
simulations or games. The algorithm determines the rules based on continuous feedback on
the actions it takes during the training.
Example: This methodology was used to train an algorithm to play Go.8
The computer played the
game against itself multiple times and gained feedback based only on the final score of each
game. Even though the algorithm was never taught about Go strategies, it was, based on this
feedback, able to improve its skills and outperform human players.
Generative Adversarial Networks
A special type of machine learning systems is generative
adversarial networks*, which refer to the interplay
between two models: a discriminative and a generative
one.
The discriminative model is trained to detect whether a
piece of data is part of a real data set or was generated
by an algorithm*, whereas the generative model is
trained to create outputs imitating those of a real data
set. Thereby, the outputs of one model serve as the
inputs for the training of the other one. The reaction of
the latter to these inputs serves as a feedback for the former. The loss function* of the
generative model hence depends on its capacity to mislead the discriminative model,
whereas the loss function of the discriminative model depends on how well it can identify
whether the inputs were created by the generative model.
Importance of Data
In general, training data is the most valuable element of the machine learning process. The
better the training data in terms of quantity, quality and variety, the more accurate the
calculation of the trainable parameters and hence the more precise the output.
Training data is not stored within the model during the training process, and once the
training process is completed, the model is fully usable independently of the data. Therefore,
the developer of a model can commercialise this model without the need to disclose the
data used for training.
Generative
model
Discriminative
model
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
9
Machine Learning Outputs
What are the possible outputs of a machine learning
model, and to what extent are they explainable and
interpretable?
Need for Differentiation between ‘Direct Output’ and ‘Application’
A distinction should be made between the direct output of a model* and the potential
practical applications of this output. Such distinction might appear formal but is nevertheless
necessary: in particular, from an intellectual property law perspective, distinct legal issues
can arise with regard to direct outputs and their applications.
Direct Output
The final output of a machine learning* model depends on what the model is trained for. It
can be e.g. a correlation between data points, clustering (e.g. defining groups of customers
based on their characteristics), or a prediction (e.g. the probability that a cat is in the picture).
The accuracy of the output depends on the quality of the model and, thus, on model
architecture*, training algorithm* and training data* (see Q3, Q4).
If a trained machine learning model is confronted with the same (new) data, it will produce
the same output. Even though the output is in principle deterministic and traceable, it is
often not human-explainable due to the complexity of the calculations, especially in the
case of artificial neural networks (the ‘black box’ issue).
Application
The output can then be put into practical use. Examples of applications might fall within the
areas of art (e.g. paintings, text conception and translation, music creation) and technology
(for instance the functioning of a self-driving car, optimisation of a car design, development
of medical treatments, virtual assistants). For instance, the goal of an application in technical
fields can be to develop a new product (such as an antenna), or to guarantee the correct
functioning of a device (such as a self-driving car). Further, machine learning outputs can be
used to improve business processes (such as targeted marketing).
Technically, an application may take different forms: the output of the model might need to
be decoded or not, and be used as such or combined with other technical or knowledgebased
elements.
Any application requires human input: a machine learning model will not independently start
using its output to do something. However, the required degree of human contribution can
vary considerably from one application to another.
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
10
Reverse Engineering of Machine Learning
Components
Is it possible to reverse engineer different components
of the machine learning process?
Reverse Engineering
In the machine learning* context, reverse engineering can be defined as the possibility to
extract or deduce certain elements of the machine learning process through access to other
elements (e.g. to deduce the model* itself by analysing predictions made by it, or to deduce
the training data* from the model).
The possibility to reverse engineer machine learning elements is controversial. Research is
currently conducted in this field. Its purpose appears to be primarily to explain the machine
learning output, i.e. to understand the factors leading a given model to a concrete output.
Currently, complete reverse-engineerability of all the parts of the machine learning system
does not seem realistic. Some elements could potentially be reverse-engineered,9
but since
there are often too many ‘moving parts’ (see Q4, Q5), this seems very challenging. The
possibility to reverse engineer varies from case to case and depends on the complexity of a
model and on how much knowledge one already has about the training process. For example,
knowing that a linear model was used, and having access to the model function (but not the
parameters) makes it easier to reverse engineer the model.
In some cases, reverse engineering might not be viable from an economic point of view, as
it might be easier to design an alternative model that produces a comparable outcome.
Teacher-Student Learning
The factual exclusivity of a model could potentially be circumvented by developing a socalled
student model*. A student model is trained to reproduce the outputs of the teacher
model when confronted with the same input. The model will therefore not be the same, but
could produce similar outcomes.
Machine Learning ‘Intelligence’
Which parts of the machine learning process are predetermined
by humans?
The human input in machine learning* mainly subsists in choosing or developing a training
algorithm* (can require creativity to develop a new algorithm), setting the hyperparameters*
(often involves trial and error; research is conducted to use machine learning to define
hyperparameters), data labelling (mundane work), and developing the model architecture*
(often a heuristic* process, see Q4, Q5).
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
11
Even if models* appear ‘intelligent’, they generate output by merely relying on probability
calculations (see Q5). They are not autonomous (i.e., they do not ‘reason’ on their own) and
need to be fine-tuned by machine learning experts.
To what extent is a machine learning model a ‘black box’
that cannot be explained by a human?
Artificial intelligence in general and machine learning* in particular are often described as a
‘black box’, i.e. even though experts can in general explain how a model* functions, they
cannot explain precisely why it generated a concrete output based on a given input. Research
is being conducted in this area (the ‘Explainable AI’ movement).
In general, the explainability of machine learning varies depending on the complexity of a
model and the training techniques used (see Q4, Q5). The problem usually arises with regard
to deep neural networks*, as unlike computers, humans are not capable of processing such
large amounts of data.
Explainability and interpretability of machine learning is currently one of the major research
areas.
3. Evolutionary Algorithms
What are evolutionary algorithms? Do they differ from
machine learning?
An evolutionary algorithm* is an optimisation method that attempts to identify the best
solution for a given problem out of multiple, autonomously generated alternatives.
Evolutionary algorithms rely on Darwinian principles, as natural evolution has proven to be
a powerful optimisation process.
In order to find a solution, an evolutionary algorithm does not need training data*, as in the
case of artificial neural networks* (see Q5). Instead, an initial population of possible solutions
with different characteristics is first generated randomly. Second, the evolutionary algorithm
evaluates the quality and/or fitness of each solution in that population and selects the bestsuited
ones. Then the selected solutions are modified using mechanisms such as
reproduction, mutation, and recombination. This process generates a new population which
is again evaluated. The process continues until an optimal solution is found.
Evolutionary algorithms can be used in machine learning*, e.g. to find the best model*.
However, their scope of application goes beyond the creation of models, since they can be
used for other tasks (e.g. an evolutionary algorithm was used by NASA for the development
of an antenna).10
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
12
Glossary
Algorithm: a step-by-step instruction. In the machine learning context, an algorithm is an
instruction coded as software and directed at a computer. (Q3)
Architecture: a structure of a model usually established by a programmer prior to the training
process. A type of architecture are artificial neural networks. (Q4)
Artificial intelligence (AI): a catch-all term that describes a branch of computer science
dealing with the development of systems that behave in a way similar to human intelligence.
(Q1)
Artificial neural network (ANN): a type of a model, the structure of which consists of layers
of neurons connected by weights. (Q4)
Deep learning (DL): a type of machine learning based on the use of a more complex
architecture (composed of a higher number of layers), often referred to as deep neural
networks. (Q4)
Deep neural network (DNN): see Deep learning. (Q4)
Evolutionary algorithms: an optimisation method based on the principles of evolution trying
to identify the best solution for a given problem out of a given population. (Q10)
Generative adversarial networks (GAN): a special type of machine learning systems relying
on an interplay between the so-called discriminative and generative models. (Q5)
Heuristic (methods): an approach to problem-solving relying on experiences and intuition
rather than on a pure scientific methodology. (Q5)
Hyperparameter: a feature of a model that is fixed before the training process and does not
evolve. (Q4, Q8)
Loss function: a mathematical function which evaluates the magnitude of error of a specific
model. (Q5)
Machine learning (ML): an automated process of identifying patterns in available data and
then applying the knowledge to new data. Machine learning is currently the most commonly
used subfield of artificial intelligence. (Q1)
Model: an algorithm based upon a (nonlinear) mathematical function that generates output
based on the patterns learned from the training data in the training process. (Q4)
Neuron: a mathematical function as a part of a model which transforms inputs (the numeric
values of the upstream weights) into an output (the numeric value of the downstream
weight). (Q4)
Offline model: a model in which optimisation process and application are separated: in the
application phase, no modification of the weights takes place. (Q4)
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
13
Online model: a model in which the optimisation never ends, because even when applied in
‘the real world’, its output is used to continually modify the weights. (Q4)
Reinforcement learning: a learning process not relying on pre-existing data but rather
gathering data from simulations or games, where the algorithm ‘figures out’ the rules based
on continuous feedback on the actions it takes during the training. (Q5)
Software: digital encoding of instructions that tell a computer how to work. (Q3)
Supervised learning: the currently most common type of a learning process in which training
data is labelled to tell the model whether the prediction made during the optimisation
process was right or wrong. (Q5)
Student model: a model trained to reproduce the outputs of a teacher model when
confronted with the same input. (Q7)
Trainable parameter: an element of a machine learning model that evolves during the
training process. (Q4)
Training algorithm: an algorithm used to train the model. It seeks to minimise its loss
function by finding the best combination of the model parameters. (Q5)
Training data set (training data): a set of data used for the training of an algorithm in the
optimisation process. (Q5)
Transfer learning: the use of a model trained for a different, yet similar task as a starting
point instead of the randomly allocating the weights. (Q4)
Unsupervised learning: the type of learning in which the model is confronted with unlabelled
data and ultimately trained to identify similarities, parallels and/or differences in data. (Q5)
Weight: a trainable parameter connecting neurons in a given architecture. It is a numeric
value that is first randomly allocated and then optimised during the training process. (Q4)
Electronic copy available at: https://ssrn.com/abstract=3465577
Drexl / Hilty et al.: Technical Aspects of Artificial Intelligence
Max Planck Institute for Innovation and Competition Research Paper No. 19-13
14
Endnotes
1
For an in-depth presentation of artificial intelligence, see European Commission – Independent
High-level Expert Group on Artificial Intelligence, A definition of AI: main capabilities and disciplines,
https://ec.europa.eu/digital-single-market/en/news/definition-artificial-intelligence-main-
capabilities-and-scientific-disciplines.
2
In rule-based systems, the reactions are pre-programmed and not adaptive (e.g. rule-based chatbots).
3
Robotics is defined as the science of designing and operating robots, i.e. machines that are controlled
by a computer that are used to perform jobs automatically (Cambridge dictionary).
4
See e.g. WIPO, Technology Trends 2019: Artificial Intelligence, https://www.wipo.int/edocs/pubdocs/
en/wipo_pub_1055.pdf, p. 26.
5
European Commission, Artificial Intelligence for Europe, COM(2018)237 final, p. 10.
6
Another exemplary breakthrough is the development of word embeddings; see Mikolov/Sutskever/
Chen/Corrado/Dean, Distributed Representations of Words and Phrases and their Compositionality,
https://papers.nips.cc/paper/5021-distributed-representations-of-words-and-phrases-and-their-
compositionality.pdf.
7
Economic literature argues that breakthroughs in relation to machine learning techniques and hence
the technological side are the reason for widespread application and technology diffusion. See
Furman/Seamans, AI and the Economy, https://ssrn.com/abstract=3186591, p. 4: ‘performance
increases are due to breakthroughs in various machine learning techniques’ or ‘these scientific
breakthroughs are starting to find their way to commercial applications’.
8
Silver/Huang/Maddison et al., Mastering the game of Go with deep neural networks and tree search,
https://www.nature.com/articles/nature16961.pdf.
9
See, for instance, Oh/Augustin/Schiele/Fritz, Towards Reverse-Engineering Black-Box Neural
Networks, https://arxiv.org/pdf/1711.01768.pdf; Tramèr/Zhang/Juels/Reiter/Ristenpart, Stealing
Machine Learning Models via Prediction APIs, https://arxiv.org/pdf/1609.02943.pdf;
Veale/Binns/Edwards, Algorithms that Remember: Model Inversion Attacks and Data Protection Law,
https://ssrn.com/abstract=3212755.
10
Hornby/Globus/Linden/Lohn, Automated Antenna Design with Evolutionary Algorithms,
https://ti.arc.nasa.gov/m/pub-archive/1244h/1244%20(Hornby).pdf.
Image sources (p. 5)
Christies, Edmond de Belamy, https://www.christies.com/img/LotImages/2018/NYR/2018_NYR_
16388_0363_000(edmond_de_belamy_from_la_famille_de_belamy).jpg.
MIT Technology Review, DeepMind has made a prototype product that can diagnose eye diseases,
https://www.technologyreview.com/f/613249/deepmind-has-made-a-prototype-product-that-can-
diagnose-eye-diseases/.
Pette, Where VR Meets the Road: How GPUs Power ‘Hack Rod’, World’s First AI-Generated Car,
https://blogs.nvidia.com/blog/2016/07/26/hack-rod-car-ai/.
Electronic copy available at: https://ssrn.com/abstract=3465577