1 What is AI?
Artificial intelligence systems perform actions, physically or digitally, based on interpreting and processing structured or unstructured data, to achieve a given goal.
«Doing nothing with AI», Emanuel Gollob (AT) – Photo: Ars Electronica
Definitions of artificial intelligence (AI) vary considerably, and often change in line with what is technologically possible. This strategy takes the definition proposed by the European Commission's High-Level Expert Group on Artificial Intelligence5 as its starting point, and defines AI as:
AI systems act in the physical or digital dimension by perceiving their environment, processing and interpreting information and deciding the best action(s) to take to achieve the given goal. Some AI systems adapt their behaviour by analysing how the environment is affected by their previous actions.
As a scientific discipline, artificial intelligence embraces various approaches and technologies, such as machine learning (including, for example, deep learning and reinforcement learning), machine reasoning (including planning, search and optimisation), and certain methodologies in robotics (such as control, sensors, and integration with other technologies in cyber physical systems).
Figure 1: Simplified overview of AI's sub-disciplines - Source: Independent High-Level Expert Group on Artificial Intelligence set up by the European Commission (2019): A definition of AI: Main capabilities and disciplines.
'Strong' and 'weak' artificial intelligence
We are still a long way from a form of artificial intelligence that resembles human intelligence, or artificial general intelligence (AGI). Artificial general intelligence is often referred to as 'strong AI' while other forms are referred to as 'weak AI' or 'narrow AI'. This does not mean that AI systems that are designed for a specific 'narrow' area cannot be powerful or effective, but they more often refer to specific systems designed to perform a single task, such as image processing or pattern recognition, for specific purposes. Nor is it the case that AI developed in parallel in many specific areas, or research on 'weak AI', necessarily brings us closer to artificial general intelligence.
Our definition embraces both 'strong' and 'weak' artificial intelligence.
Rule-based systems for automation
A rule-based IT system is often built on rules such as 'IF x happens, THEN do Y'. Such rules can be organised in complex decision trees. Rule-based automation systems can be used to model regulations, business logic or experience-based practice (exercise of discretion). Many of the systems used for automated administrative processing in the public sector are rule-based. Our definition of artificial intelligence covers some of these systems, depending on factors such as the complexity of the rule set.
1.2 How does artificial intelligence work?
A system based on artificial intelligence can either interpret data from devices such as sensors, cameras, microphones or pressure gauges or can be fed input data from other information sources. The system analyses the data, makes decisions and performs actions. Both the need for data and the fact that it is the system that makes decisions and performs actions raise ethical issues that are discussed in chapter 5. Some types of systems have a feedback loop which enables the artificial intelligence to learn either from its own experiences or from direct feedback from users or operators.
The artificial intelligence system is usually embedded as a component within a larger system. Tasks are often performed digitally, as part of an IT system, but AI systems can also be part of a physical solution, such as a robot.
Examples of current practical applications of AI are:
- Computer vision/ identification of objects in images: can be used for purposes such as facial recognition or for identifying cancerous tumours.
- Pattern recognition or anomaly detection: can be used to, for example, expose bank or insurance fraud or to detect data security breaches.
- Natural language processing (NLP): can be used to sort and categorise documents and information, and to extract relevant elements from vast datasets.
- Robotics: can be used to develop autonomous vehicles such as cars, ships and drones.
Development in some areas has progressed rapidly, and we are already seeing systems being used in practice. Development and testing in other areas can take longer to achieve reliable and verified results.
Today when we hear about systems being based on artificial intelligence, they are usually based on machine learning. Unlike rule-based systems, where rules are defined by humans and are often based on expert experience, business logic or regulations, the concept of machine learning covers a range of different technologies where the rules are deduced from the data on which the system is trained.
In AI systems developed by machine learning, the machine learning algorithms build mathematical models based on example data or training data. These models are then used to make decisions.
Machine learning algorithms usually learn in three different ways:
- Supervised learning: the algorithm is trained with a dataset where both input data and output data are given. In other words, the algorithm is fed both the 'task' and the 'solution' and uses them to build the model. This will make it capable of making a decision based on input data.
- Non-supervised learning: the algorithm is fed only a dataset without a 'solution' and must find patterns in the dataset which then can be used to make decisions about new input data. Deep learning algorithms can be trained using non-supervised learning.
- Reinforcement learning: the algorithm builds its model based on non-supervised learning but receives feedback from the user or operator on whether the decision it proposes is good or bad. The feedback is fed into the system and contributes to improving the model.
Figure 2: The interrelationship between an AI system, its operator and environments.
Deep learning is a subcategory of machine learning. Today deep learning is an important component in widely used solutions such as image processing, computer vision, speech recognition and natural language processing. Other areas of application are: pharmaceutical development, recommendation systems (for music, films, etc.), medical imaging processing, personalised medicine, and anomaly detection in a range of areas. The most widely used deep learning frameworks have been developed by Google (TensorFlow) and Facebook (PyTorch).
Some deep learning algorithms are like a 'black box', where one has no access to the model that can explain why a given input value produces a given outcome. This is discussed in more detail in chapter 5.