The hype about the possibilities and possible applications of artificial intelligence (AI) seems currently unlimited. The AI procedures and solutions are praised as true panaceas. However, when viewed soberly, they are just another tool in the toolbox of IT experts.
There are scenarios in which AI applications deliver better results, but no general superiority can be derived from this. More importantly, IT managers have to check very carefully what they want to use for each project. To answer this question, decision-makers must consider AI in connection with other concepts.
AI does not replace, AI supplements
Two different procedures can be distinguished in IT: the programmatic or the model-driven approach. The first is the traditional development of software. Programmers write code line by line, specifying which requirements a system should meet and how: from the collection of address information in a CRM database to the calculation of tariff options in a health insurance scheme to the speed of a milling machine depending on the workpiece. Typical representatives of this approach are standard solutions such as Enterprise Resource Planning (ERP) software, but also specific software or mixed forms that combine standard with individually created components.
The model-driven approach, on the other hand, involves less programming — in the sense of writing down lines of code — but focuses more on technical, functional aspects. A stronger abstraction of technical aspects accompanies this. It is about the illustration of correlations and effects in a model. This is modeled or trained manually, and the knowledge gained can then be applied to new input, such as other data sets.
In principle, both IT solutions are of equivalent value concerning the result — both a programmed system and an AI can solve the same task. However, the selection of the solution has an impact on the project design, on the parties involved, on the effort and finally on the entire life cycle of the IT solution.
AI projects are similar but different.
If the responsible persons rely on a software development process, the project goes through the classic steps of planning and implementation: Understanding requirements, developing, testing, rolling out and operating software on this basis. New or changed conditions ensure that those involved go through this cycle again and again.
The procedure for a model-based approach is different: In this case, everything starts with understanding the requirements as well. However, then it is necessary to pour these requirements into a model, to check the quality of the model with test data, for example, to integrate it into IT operations, roll it out and operate it. The typical IT implementation project takes place in the same way as the process described above, i.e., after that. However, the functionality regarding the required functionality is already given by the model. The implementation project then “only” deals with so-called non-functional requirements (NFR) such as performance, scalability or availability. This procedure, therefore, ensures clear separation of functional and technical concerns. Accordingly, in comparison to programming, the technical experts and modeling have a significant role to play in this procedure; IT competence is less relevant for this.
There are two different approaches to modeling: On the one hand, the experts can describe the model manually, for example by writing down the rules. On the other hand, they can rely on self-learning systems (machine learning). Here, a learning algorithm replaces human developers. For this algorithm to work, it needs input. Either as training based on historical data sets — for example, categorized photos — or as direct feedback from the experts. In which situations should decision-makers choose which method? This cannot be answered in general terms, but some indicators indicate trends.
No fixed rules for AI use
As already explained, AI does not replace anything known but complements the IT solution space with further options. The following indicators support the choice of an AI-based solution:
– The technical complexity is high and worthy of a model.
– The separation of business issues and technical necessities brings advantages. For example, because business experts can expand and optimize the system independently.
– The experts are only able to map and maintain the technical aspects with a great deal of effort. In these cases, the use of learning systems can pay off.
So much for theory. An example illustrates how programming or modeling by hand, in contrast to machine learning, proves itself in the same application case.
The scene for this comparison is the naturally somewhat restricted area of military surveillance: these are classified for the analysis of radio signals. As long as they are not encrypted, the contents are then decoded. The material can be either voice or data. In the case of speech signals, the experts distinguish between analog speech and speech signals that are digitally coded. It may be surprising considering the technological developments, but there are still many analog, uncoded signals in use.
Listening to these radio signals requires many people to capture and transcribe analog radio signals every day. A time-consuming and expensive process. The process would be much simpler if a system could automatically distinguish analog speech signals from digital signals.
To create the basis for such a system, the radio signals involved in the project were first manually classified into the categories analog or digital. They listened to the radio signals and marked the beginning and end of each analog speech signal. In this way, they marked radio signals with a duration of approximately 24 hours. The project managers then used two different methods: In the first approach, a team programmed algorithms to extract characteristics from these signals. These characteristics were then classified by programmed case distinctions — an expensive, time-consuming project.
A deep neural network takes over the signal evaluation
In contrast, a master thesis dealt with a solution based on self-learning processes. The student used a so-called deep neural network, which was trained with the classified data. The student left both the selection and the evaluation of the signals to the neural network. So nothing was programmed; the preparation consisted of the formatting of the data material and the configuration of the network. A functional system was already available after about three months of operation.
A comparison of the two methods shows that the learning system achieves better results despite the low effort involved. It produced a recognition rate of 95 percent for analog speech signals — well above the rate achieved by the programmed procedure.
The responsible persons must therefore carefully consider in each case if the use of AI technologies is the best option. It is rarely possible, as in this example, to develop several procedures in parallel. Human intelligence and experience are then required to select the appropriate method.