Not a Mystery Any Longer: Artificial Intelligence

Artificial Intelligence … a much-discussed buzzword that has become an integral part of almost every industry, but how does it actually affect the printing industry and how does it work? While we have already explored the first question in various articles – from defect detection in 3D printing software to increasing efficiency in recycling facilities to the use of AI in the packaging world – we have so far always left out the rules and operating principles behind it. That is to change now.

The Basics of Artificial Intelligence

Let’s start from the beginning. Artificial Intelligence is a branch of computer science dealing with the automation of intelligent behaviour and machine learning. It has the goal to develop cognitive systems that solve problems independently by mimicking human intelligence. It is important that they themselves are able to interpret influences and react to them intuitively – to do this, a computer evaluates data and draws conclusions, respectively learns from mistakes and perfects its behaviour automatically.

This is usually made a reality by so-called artificial neural networks (there also are heuristic methods and evolutionary genetic algorithms), which copy biological and evolutionary mechanisms and mimic the functioning and information processing of the brain in order to find the best results from a set of solutions based on algorithms. An algorithm, for its part, is an exact step-by-step recipe or sequence of different rules that make it possible to solve tasks.

Self-learning Systems

AI can thus be applied wherever solutions can be generated by comparison with similar problems or by recognising patterns. The ideal conditions for these self-learning systems are therefore places where plenty of material for learning and recognising patterns is available.

Although we are dealing with computer science here, these sequences of explicit instructions do not necessarily have to be recorded in a programming language, but are often even simply written down in natural language. Still, the ultimate goal is always to put the algorithm into machine-readable form. This includes writing a computer programme based on the processing instruction that solves the problem in question.

About Decision Trees, KNNs and Linear Regression

Now, we know what role algorithms play in AI systems, but how do they actually function? To understand what exactly happens when artificial intelligence works its supposed magic, you don’t have to be a technology expert, as we will explain the process using the three AI methods “decision trees“, “K Nearest Neighbours” (kNN) and “linear regression” as examples. The first two belong to the so-called classification algorithms, while linear regression belongs to the regression algorithms. Classification algorithms are used to divide an examined variable into different classes and then predict the class for a given input. For instance, one application of classification algorithms is to classify emails as spam or not. Regression algorithms can predict output values based on input data points fed into the learning system and are mainly used to forecast stock prices or the weather. Additionally, there is also the algorithm category clustering, which describes a process of segregation and organisation of data points into groups according to similarities within the group members. In contrast to the two algorithm categories that we will present in more detail below, it does not belong to supervised learning, but to unsupervised learning. As you can see, there are many more algorithm types, but our selection is a good starting point to open up the topic initially: 

1. Decision Trees
   … belong to the simplest algorithms and are inspired by the branches of a tree crown. They resemble a flowchart and many will certainly know their principle from other areas of life. Starting with a question, the answer decides which question follows, until at the end of each path there is an assignment to a class. Accordingly, the nodes represent the test for an input attribute and the branches lead to the outcome of the test. An example of this is a programme that is supposed to make a credit decision. To do this, it asks for data about the client in question, e.g. about his work situation. If the customer has been employed for less than three months, a different question is asked than for a customer who has been working in a present job for more than three months, until the programme has finally collected enough information to either accept or reject the credit application. In the field of artificial intelligence, however, this process is not pre-programmed by a software developer, but instead only determined by the algorithm that automatically generates such tree structures from a given data set with reliable empirical values. It then determines which questions are to be asked and how they are to be arranged depending on the answers.
2. K Nearest Neighbours
   … describes a procedure that can make predictions about inspected data points on the basis of information already collected about similar data points. These are then assigned to the class to which most of the neighbours are also assigned. Here, k stands for the number of neighbours included, so that different results can arise depending on the size of this parameter. For example, if 1) a new data point is added to a system, 2) k is set to 5 and 3) 4 of its 5 nearest neighbour data points are small, then it is automatically classified as small. If, on the other hand, k is set to 10 and 6 of its nearest neighbours are large, then it is classified as large instead of small. Accordingly, it is important to carefully consider the optimal definition of k depending on the respective application and also to define the concept of neighbourhood sensibly.
3. Linear Regression
   … is the simplest of all regression algorithms, in which the regression or best fit line is drawn between data points to predict new values. This model is applied when a linear relationship between individual parameters is suspected and is mainly used when estimating real values, e.g. the cost of a house based on its square metres. This can be imagined in the way that purchase prices for houses are entered into a diagram depending on their size. With a line as close as possible to all points, the best linear fitting relationship is mapped. In this way, a known value such as the number of square metres can be used to draw conclusions about the second data point – in our example, the value.

We hope this article and especially our examples shed some light on the various Artificial Intelligence algorithms and could show how simple the principle behind them often is. In the end, however, it is not so important how exactly such algorithms work, but it is rather essential to know which are suitable for which applications.

Have you ever had a closer look at artificial intelligence systems and do you already use them in your everyday work? Tell us about your experience in the comments section.