Artificial intelligence is a new technology

Artificial Intelligence

From Alexa to autonomous driving, from algorithms to omnipotence fantasies - the spectrum of topics associated with artificial intelligence (AI) is wide. Experts and laypeople alike bring very different processes and concepts into the field under the heading of AI. In fact, each approach has different possibilities and is suitable for special application scenarios in companies.

In general, AI is a branch of computer science and deals with the mapping of "intelligent" behavior through IT. AI has not made the leap from universities to companies for a long time. The technology was unable to meet many of the expectations of the early years. But now the tide is turning: the media keep reporting about breakthroughs, new areas of application and new potential dangers.

Three factors are responsible for the triumphant advance of AI: more data, cheaper storage capacities and constantly increasing computing power (for example through NVIDIA graphics card farms). They enable companies to use AI processes in increasingly complex configurations.

Experts differentiate between "strong AI", the aim of which is to imitate human intelligence, and "weak AI", which is used to make intelligent decisions for specific areas, such as the automation of processes. Strong AI is beyond the current technical possibilities.

Unresolved fundamental problems ensure that it will remain a mental game of theorists for the foreseeable future - even if the reporting sometimes suggests otherwise. Weak AI, on the other hand, is an approach that plays a role in many applications today. In the following, we therefore only want to deal with aspects of weak AI.

The core of an AI system is a model

Artificial intelligence encompasses an extensive set of methods, processes and technologies. The core of an AI system is a so-called model that is modeled for a specific question - for example to support certain decisions or to make predictions.

There are many different types of models as well as different techniques for creating models. The easiest way to explain which ones are used is based on specific use cases. More on that later.

Basically, AI can be divided into so-called symbolic and sub-symbolic systems based on the representation of knowledge: In a symbolic system, rules and relationships are used for concepts that are understandable to humans. That means: The model can be read and recorded by people. Sub-symbolic systems, on the other hand, are largely black box systems for people, the contents of which are not easy to understand.

Symbolic and sub-symbolic system

The following example illustrates the different concepts. The aim is to identify the fathers in a group of people. In a symbolic system, experts state that the concept "father" is a specialization of the concept "person"; namely a person who is male and who has at least one parenting relationship with another person (his child).

If such rules are stored, the system can answer search queries for fathers - even if the "father" property is not specified in the searched data, but only information about the relationships between people and what gender people have.

In a sub-symbolic system, however, experts could use a so-called support vector machine (SVM) or an artificial neural network (ANN) for the same task. Both are procedures that experts use frequently and successfully for the automatic classification (classification) of data for certain classes. In principle, these methods are also suitable for classifying people as fathers and non-fathers.

For this purpose, the support vector machine divides a multi-dimensional data space in the form of a level - a so-called hyperplane - that separates the groups of fathers and non-fathers. In a simplified example with two dimensions, this means that points in a coordinate system represent the data to be evaluated and the SVM uses these points to determine a straight line that separates the points into two classes (father / non-father).

The neural network solves the "father / non-father" classification entirely differently, namely by deriving these classifications from a set of networked "neurons". These map various data transformations. The neurons are each grouped into levels, and neurons on neighboring levels are linked to one another by weighted connections. Thus, data entered into the network are processed by the neurons of the various levels in order to finally activate the "father" or "non-father" class.

Hidden Markov model

In addition to the support vector machine and artificial neural networks described, there are other methods for modeling. The so-called Hidden Markov Model (HMM) is known, which plays a role particularly in the area of ​​predictive maintenance.

  • The HMM is based on the so-called Markov property: A data point only depends on the data point before (often also on X data points before: Markov model of the Xth order).

  • The model learns to generate short time windows for the data from the collected data.

  • If the model is very poor at predicting a time window, this signals an anomaly and indicates a problem.

Using the example of the evaluation of sensor data on a motor, the functionality of the HMM can be explained: The HMM learns the relationship between the number of revolutions and temperature over time. The model continuously predicts the number of revolutions and temperature data, these are compared with the real measured data. If the real and the predicted data deviate significantly from each other, the experts assume a defect or a problem that is just becoming apparent.

These explanations alone make it clear: The process and the functioning of sub-symbolic systems are difficult to convey in just a few words.

In the symbolic system, people can easily understand how the "father / non-father" decision-making process works: they just have to read the rules and apply them. The system can even "explain" the decision as to why a person is a father by showing how what information contributed to the decision.

The traceability of the decision in favor of a sub-symbolic system - in the example mentioned, a support vector machine or a neural network - is hardly comprehensible. Understanding the process that led to the creation of levels in multidimensional spaces or the weighting of relationships between neurons is extremely complex.

Back to business practice: which system is suitable depends on the context of the application and the company. If the regulations require that decisions about, for example, loan commitments or the approval of construction procedures must be traceable and transparent, sub-symbolic procedures for companies or authorities are ruled out. They do not offer any means of providing such explanations.

Use cases for AI systems

No intelligence without learning or modeling - this also applies to artificial intelligence. So-called machine learning (ML) is the ability to automatically learn a model based on data (see also the article: What is TensorFlow?). Machine learning is available in the forms of supervised, unsupervised and reinforcement learning:

  • Supervised learning means that experts have to tell the process what the right decision is in each case for a set of training data. Since large amounts of training are often required to get good results, the effort involved is often high.

  • In the case of unsupervised learning, the system analyzes data with regard to their similarity or their distance without experts entering training data. An example of unsupervised learning is looking for classes in a set of data points. Usually the only input in unsupervised learning is the number of classes to be found.

  • Experts use the term reinforcement learning to summarize processes that learn in the form of direct feedback - but not by providing training examples. Well-known application scenarios for this form of learning are learning to play games such as chess, Go or various computer games.

Regardless of the method chosen, a model can identify relationships and provide new insights. This ability gives rise to a wide range of possible uses:

  • Early warning systems in mechanical engineering in which the system learns to interpret the mechanisms that lead to machine failure at an early stage (predictive maintenance).

  • The partially or fully automatic recognition of relevant text passages in unstructured documents, for example in the process of reporting damage in insurance companies.

  • Automatic procedures for discovering and preventing acts of fraud in the finance or insurance industry, the so-called fraud detection.

The models can now even be used as an autonomous cameraman. The Essen start-up offers a solution that takes over the camera work during a soccer game. Here an AI model was successfully trained that can independently select relevant camera positions and settings.

The right solution emerges from the right combination

Complex systems usually use different AI methods. The following use case from the banking environment shows how methods interact.

The fund managers in banks rely on in-house advisors for their investment decisions. These experts read and evaluate analyzes, for example on the development of industrial sectors in different regions of the world. However, it is almost impossible for individuals or teams to keep an eye on all the relevant sources of information and then to be able to access the relevant knowledge about the relationships. AI processes make it possible to analyze reports automatically and then make the content available in the form of natural speech output.

With the help of methods of natural language processing, texts are analyzed and information is extracted that can be processed by machines, that is: that can be "understood" by machines. In contrast to most people, a text simply represents a set of characters without meaning (semantics) for a machine. The information obtained in this way is fed into a system. This contains a complex model that can suggest investment decisions.

The financial experts can send the request for an investment recommendation in the form of natural language: "Please give me an investment recommendation with a focus on Asian markets in the 'income-oriented' risk class." The voice signal is then translated into text using the same processes that systems such as Amazon Alexa, Google Home, Apple Siri or Microsoft Cortana use. This is mostly machine learning based on neural networks. In the text recorded in this way, the system then recognizes the aim of the questioner, the so-called intent. The knowledge representation system generates the investment recommendation from this intent, which is then read out.

In this use case alone, methods of natural language processing, knowledge representation and machine learning work together to provide the desired functions.

Differentiate clearly when planning

The examples and explanations show that there are a number of different methods, processes and technologies behind the topic of AI. In order for a project to achieve the desired success and meet all regulatory requirements, the parties involved must differentiate clearly during planning. It is crucial to select individual components in such a way that they all meet the requirements for transparency and traceability. Often there is more than one path to the goal - but not all paths are open to a company.