Artificial intelligence – an introduction

“Machine beats chess world champion”, ‘Computer flies drone’, ‘In the board game Go, humans have no chance against artificial intelligence’. These or similar reports lead to the assumption that computers and machines could soon outstrip humans with their intelligence. Science fiction scenarios fuel the fear of a loss of power, the Frankenstein complex threatens to become reality: a machine turns against its creator, the matrix takes power and we humans are the livestock of an all-powerful and omniscient intelligence.

Artificial intelligence – a definition

This sometimes irrational discussion makes it difficult to take a realistic view of artificial intelligence, and so there is no generally accepted definition. What we don’t understand can be frightening. I would like to counter this fear, which is not a good advisor, with a small series of blog articles and, on the contrary, inspire joy in dealing with AI topics. Artificial intelligence, used on the good side of the force, is already generating invaluable results, for example in the detection of cancer, and could contribute to immense value enhancement in a wide range of industries and for a wide range of business ideas, particularly in Germany.

First, let’s try to define it:

• For me, artificial intelligence is an autonomous and adaptive system that performs complex tasks in complex environments without being aware of itself.
• Autonomy in this context means the ability to perform tasks in a complex environment without permanent user guidance (i.e. the machine is not programmed by a human in advance). Adaptive means the ability to learn from experience and thereby improve one’s own performance.

The computer scientist Tom Mitchell provides a more formal definition: “A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.”

How does machine learning work?

In addition to a definition that provides a first idea of what AI does, the following aspect is fundamental to understanding what AI can do today:

In reality, every system is modeled for a specific purpose and made to work for that purpose through hard work. To illustrate this:

• Text-generating systems cannot assess credit risk.
• Fraud detection software cannot drive a car.
• House price forecasting software cannot detect breast cancer.

So how does machine learning work in all these cases, the results of which are impressive in each individual case?

There are basically three types of machine learning:

Supervised learning

Supervised learning can be thought of as learning based on many examples. The student here is a mathematical function that is often highly complex or a neural network. A neural network is also a mathematical model that offers highly complex mathematical modeling options. The student learns by gradually looking at a huge amount of data as examples of the answers that will be required of it in the future. A simple example: The student is given many data sets of the form “house size: 110 m2 price: 250000 €” and generalizes his abilities with a large number of learning examples. If he is then asked to name a price for a house size of 105 m2, he can answer with a very good approximation.

In the example just described, the machine predicts a number, the house price. This type of machine learning is called regression, and is also generally referred to as a regression problem. A second type of supervised learning is called classification. In classification problems, the machine can assign an input data set to a class after training with sample data. Is the image of breast cancer, yes or no? Is the handwritten number 1 or 2 or another number or even a letter? Is the e-mail spam or not?

Unsupervised Learning

Unsupervised learning is also based on training the machine with a lot of input data. Unlike supervised learning, however, the data sets do not contain the answers regarding class or a number. With the mathematical models used in unsupervised learning, the machine learns independently, without the input of answers, to recognize structures in the data. Examples of where unsupervised learning is used include assigning customers to segments, recognizing relationships in a business intelligence database, or recognizing the structure of anomalies in the production of workpieces.

In some cases, it is even possible to use unsupervised learning and supervised learning one after the other. This methodology is called pipelining. An unsupervised learning algorithm recognizes a structure in a data set, especially where outliers are located in the data space. With this knowledge, a supervised learning algorithm can then be trained (because it now has an answer in each data set, the assignment). Such techniques are used, for example, in fraud detection and in the detection of anomalies.

Reinforcement Learning

In reinforcement learning, on the other hand, the student, known as the agent, is not given any data in advance for training. It is able to perform certain actions. As an agent that wants to learn how to play chess, for example, it can move pieces around on the chessboard. The student now learns by performing actions and then receiving a reward from his so-called environment for each action. The reward depends on the learning objective (e.g. checkmating the opponent) and indicates whether the action was “good” or “bad”. In the game of chess, the environment is determined by the opponent and the board with the pieces. With each action performed, the student learns a little more about how to behave in a given environment. Reinforcement learning is used in practice for control problems, in games (the GO world champion lost to a reinforcement agent) or in autonomous driving.

I hope this brief summary has given you an introduction to the basic concepts of machine learning. Each of the different types of learning has various types of algorithms, which in turn offer solutions for certain types of problems. This basic understanding is essential for managers to be able to start a conversation with experts on an equal footing. I hope you enjoy learning more about this topic and look forward to your suggestions and comments.