Artificial Intelligence: Smart and Self-learning

Author: Markus Strehlitz

Nov 10, 2021 Artificial Intelligence

Artificial intelligence does amazing things in many different areas – from quality control in manufacturing to diagnosing skin cancer. Testing and certification is an absolute prerequisite for low-risk operation of relevant systems. And a company like DEKRA plays a key role.

Talking about artificial intelligence (AI) may sound a bit mysterious at first. After all, the term encompasses technologies which ensure that machines can solve human-like tasks. At least, that’s one simplified definition. However, a closer look reveals that current AI systems don’t have any intellectual capabilities. Such so-called strong AI is still a thing of the future.
Weak AI, on the other hand, is increasingly taking over our everyday lives. These are primarily self-learning processes that are geared to precisely one task. The systems aren’t programmed like classic software but trained with a large amount of data. When in use, they’re constantly learning.
Chatbots take over communication with customers
Such systems perform amazing things in many different applications. Intelligent chatbots, for example, take over communication with customers for companies. When people complain or want to exchange a product, they don’t speak to a flesh-and-blood person but to a pre-trained computer program.
In manufacturing companies, AI systems analyze data gathered from manufacturing machines to detect problems at an early stage. As an example, BMW collects sensor data from 600 welding guns used by robots in the body shop at its Munich plant. Software continuously evaluates the data with the help of AI and reports when failure is imminent.
Machine Learning in Quality Control
Quality control also relies on machine learning. At Fiat Chrysler China, an AI system uses cameras to identify defective assemblies or missing components, such as screws so small that they’re difficult for the human eye to detect. The parts in question are then weeded out. AI software works similarly in medicine, for example by being fed images of skin cancer variants. If the number of training images is large enough, the system can then make appropriate diagnoses on its own. In individual studies and under certain conditions, the AI was able to deliver better results than a human doctor.
Another area of application for artificial intelligence is automated driving. The more tasks the machine is supposed to take over from humans, the higher the degree of AI application. In this case, however, the application’s limits also become apparent. In automated driving, the demands on AI systems are significantly higher than in quality assurance of a manufacturing company. A camera aimed at a product for inspection generally operates under non-varying conditions. By contrast, city center traffic is very complex. Vehicles can run into unpredictable situations.
DEKRA plays a key role in AI
The field of automated driving in particular makes it clear how important it is to test AI. The relevant systems must be tested and certified in order to be utilized without risk. This applies to assistance systems in vehicles as well as to software that supports companies in the selection of new employees. The EU is currently working on a framework law for the use of AI. A first draft of the Artificial Intelligence Act was published this spring and should ideally be implemented in EU law by the end of 2022.
Dr. Tarek Besold sees a company like DEKRA in a key role when it comes to AI. He says that as a trusted party, it can support and verify compliance with common standards. “We’re already helping to create technological standards and norms for AI with our expertise from technical testing and application domains. As soon as there are regulatory defined requirements for AI systems, DEKRA will then also test and certify their compliance as a neutral independent body,” said Dr. Besold, Head of Strategic AI at DEKRA DIGITAL.
AI hub as a starting point for initiatives
In addition, DEKRA is also working on using the possibilities of AI itself – for example, to further develop existing products and services. “Many things are possible, from automatically generated inspection reports to smart camera-based inspection tools,” says Dr. Xavier Valero, Head of Applied AI at DEKRA DIGITAL.
DEKRA has established its AI Hub to drive the topic forward – across the group as well as the divisions. The hub is intended to serve as a contact point for AI initiatives within the organization, support the implementation of corresponding projects, and establish DEKRA as a participant in the national and international AI ecosystem. This way, DEKRA will also help to ensure that AI no longer seems quite so mysterious, but instead works transparently, safely, and reliably.
Three Questions for Dr. Tarek Besold, Head of the AI Hub at DEKRA DIGITAL
Mr. Besold, what criteria are used to test an AI system?
Besold: First of all, there are the classic criteria such as functional safety and cyber security, which remain relevant. In addition, it needs to be checked whether an AI system operates without discrimination. This issue is currently being discussed at the European level. After all, many AI applications automate decision-making – for example in medicine as well as in the granting of loans by a bank. It must be ensured, for example, that the results aren’t influenced by the ethnic origin, gender, or sexual orientation of the person in question.
Is the fact that AI systems are constantly learning and thus constantly evolving a challenge in testing?
Besold: It will become necessary to test AI systems on an ongoing basis. We’re already working internally on scenarios that necessitate what is known as “permanent monitoring”. This is a major challenge. When it comes to machine learning, there are no explicit rules in the systems, only statistical relationships. How a system has changed can only be judged by the result. Thus, just recognizing when the time has come to check it again is a challenge.
This brings us to the black box problem with AI. Their decision-making processes are difficult to comprehend.
Besold: That’s right. That’s why testers need both technical AI know-how as well as expertise about the application domain. In most cases, it will most likely come down to scenario-based testing – in other words, testing for the individual case in question. The nice thing about an organization like DEKRA is that we have a lot of experience with the relevant application domains.