Key concepts of artificial intelligence
Artificial Intelligence (AI), without any doubts, has been the hot topic of 2019 and it will for sure not leave us in 2020. The European Union has recognized the pivotal importance of Artificial Intelligence and has started developing an European approach to AI through a number of measures, which include the financing of program Horizon 2020 and the creation of the High-Level Expert Group on Artificial Intelligence (AI HLEG).
Thus, what seems to be even more widespread is the misuse of these two words, especially by laymen (i.e. not engineers or computer scientists) operating in the legal and medical framework. This article hopes to shed a light on some key concepts of Artificial Intelligence.
Optimistically, this will give us the basis to embark on a deeper analysis of legal aspects connected to the use of Artificial Intelligence in healthcare.
The legal aspects that will be analysed in the course of this column are various: we will talk about the relationship between Artificial Intelligence and compliance with the principles of processing of personal data enshrined in the GDPR and further regulations in the field, transparency and explainability, AI and automated decision making (art. 22 GDPR), profiles concerning AI as a medical device in light of European Regulation (Regulation 2017/745-MDR), and professional responsibility for healthcare providers when using AI.
Let’s begin by clarifying the key concepts.
One of the main issues regarding Artificial Intelligence, from a legal point of view, is that there’s not broad consensus on its definition amongst the scientific environment. Furthermore, there is a general confusion among the population on what is artificial intelligence and what is just science fiction. On this matter, the University of Helsinki, together with Reaktor, launched in 2018 a free university course online with the aim of educating 1% of European citizens on Artificial Intelligence by 2021.
As highlighted in the course, one cause of this confusion is that activities that are simple for a human being are actually difficult to reproduce through an Artificial Intelligence. If we think for example of a simple task such as looking around to find an object to collect, this is actually the result of a series of complex operations in sequence: using sight to analyze what surrounds us, finding an object that is suitable to be collected, planning the right trajectory to reach it, contracting the right muscles to make the movement and applying the right amount of force required to hold the object between your fingers. Similarly, operations that require a human being to experience and training, such as solving a complicated equation or playing a chess game, are very suitable activities for a computer, which follows certain rules and is able to test alternatives at a surprising speed.
According to the definition provided by the AI HLEG,
“Artificial intelligence (AI) systems are software (and possibly also hardware) systems designed by humans that, given a complex goal, act in the physical or digital dimension by perceiving their environment through data acquisition, interpreting the collected structured or unstructured data, reasoning on the knowledge, or processing the information, derived from this data and deciding the best action(s) to take to achieve the given goal. AI systems can either use symbolic rules or learn a numeric model, and they can also adapt their behaviour by analysing how the environment is affected by their previous actions.”
The HILEG refers to AI systems since usually Artificial Intelligence is embedded as components of larger systems. We can encounter AI-based systems as stand-alone software (such as voice assistants, image diagnostic, search engines) which operate in the virtual world and systems embedded in hardware (such as robots, drones).
AI systems are autonomous, meaning that they can perform tasks without constant guidance by a user and adaptive, meaning that they can improve their performance by learning from past experience.
The meaning of Artificial Intelligence doesn't end there. It also indicates a particular branch of computer science which, includes numerous approaches and techniques such as machine learning and robotics.
It's the fundamental unit of Artificial Intelligence. Technically, it is defined as a finite sequence of operations or instructions that from an initial set of data (input) obtains a result (output) that meets a set of pre-assigned requirements. As stated in the recent Ethical and Legal Charter of Artificial Intelligence drafted by the Leonardo Foundation, given that the algorithm is a mathematical operation and is predictable, it is precisely the definition of requirements that represents the essential aspect in which the intervention of a human operator takes place.
Consequently, it will be in the definition of the constraints and objectives of the algorithm that most ethical and legal implications will occur.
Through machine learning the algorithm is presented with example inputs of the task it should perform. In this way, human beings train the system by supplying it with data from which it will start learning. Artificial Intelligence which is based on machine learning comprises of an algorithm which is designed so that its behaviour can evolve throughout time, making decisions on how to accomplish the demanded task.
SUPERVISED/UNSUPERVISED MACHINE LEARNING
Supervised algorithms refer to the situation where an algorithm is trained based on input data which is labelled by humans. The algorithm will then define the rules of classification based on examples which are validated cases. Supervised learning implies that supervisors teach the machine the output which it has to produce. The supervisors already know what the output value should be for the samples provided to the machine for training.
An example of a supervised learning algorithms are software used for analysis of medical imaging: based on the instructions given by doctors (a certain anatomic alteration means the diagnosis of a specific disease) they are trained to classify whether a patient has a disease or not.
On the other hand, unsupervised learning implies that unlabelled input data is given to algorithms which then produces its own classification based on a pattern or a variable.
In this field, as can be seen, the quality of the data used to "train" the algorithm is of utmost importance. Any type of bias in data classification, in fact, would lead the AI to produce a discriminatory result, following the famous principle called GIGO: Garbage In, Garbage Out.
We can already catch a glimpse of the legal issues such as data ethics, the respect by a system of AI of the fundamental principles of an individual and of data protection as enshrined in the GDPR, such as the principle of data protection by design and by default. These issues, as announced, will be the subject of specific focus.
It’s a subclass of machine learning which involves algorithms that analyze data through multiple layers of complex processing. This means that the output of each layer is used as the input for the next layer. It is inspired on the processing of visual information in the brain by the retina.
ARTIFICIAL NEURAL NETWORKS
It’s a type of deep learning architecture which simulates the human brain structure in its lower-level data processing.
Although not exhaustive, we hope that this very short glossary will serve as a useful tool to understand very complex concepts such as those related to Artificial Intelligence and that it will be of support for the more in-depth analysis that will follow in the coming weeks.
 See https://ec.europa.eu/digital-single-market/en/artificial-intelligence for summary of tools adopted by the EU in the development of an EU policy in AI.
 Artificial Intelligence for Europe, Communication from the Commission to the European Parliament, the European Council, the Council, the European Economic and Social Committee and the Committee of the Regions on Artificial Intelligence for Europe, p.1. Available at: https://ec.europa.eu/newsroom/dae/document.cfm?doc_id=51625