"An Evolutionary Step"
Expert Interview with Markus Schatzl
Markus Schatzl, Director of the senswork Innovation Lab, explains what opportunities arise with the use of AI and what hurdles are associated with it. Matthias Günther conducted the interview for his thesis “Artificial Intelligence in Quality Management - Implementation Options along the Value Chain and Developments in SMEs” in the Industrial Engineering and Mechanical Engineering course at the University of Konstanz.
How do you assess the current state of AI and what progress can be expected in the next few years?
The field of application of AI is very broad, so that the question is not easy to answer. For our field of activity, it is highly relevant that the accuracy and performance are now extremely good. Due to the development progress in the hardware area, a mainframe computer is no longer necessary to process interesting tasks. One topic that is slowly arriving in the application area is how to better explain the behavior of a neural network. In the past this was not possible, so a result simply had to be accepted. This was not a problem in the manufacturing environment, as it is of course more attractive to have an automated solution with an accuracy of 99.5% than none. In the medical technology and automotive sectors in particular, the ability to explain a decision in quality assurance is central, so that this development will certainly open up further applications.
How reliable are today's methods, especially with more complex components or requirements?
That depends on the specific application. The classification of parts and object recognition in components is usually easy to solve. Certain properties can make detection more difficult, but there are effective methods in industrial image processing that can be used to improve the initial situation. This can be a special recording technique or data preprocessing with conventional image processing technology. In addition, the hardware for computing more complex networks is now very powerful. The achievable performance depends largely on the financial investment in a system.
For which company sizes are AI technologies best suited?
From my point of view, this can hardly be determined by the size of the company. The infrastructure and important tools required for an implementation are freely available. This enables a one-man / woman company in the same way as a corporation to implement AI-based applications. From my point of view, the limitations are in the area of high-performance applications, as the computing power required can be a financial issue.
Are there other minimum requirements for companies, such as technical conditions or sales, in order to implement AI?
Since computing power can be rented via the cloud, it has been possible to implement conception and preparatory work on an average workstation and then train on a high-performance, temporarily rented computing platform. In principle, the use of existing neural networks is usually unproblematic, the work processes are relatively simple. On the other hand, it is difficult to design your own complex network architectures, but this rarely happens in the application area. This difference is often not clear from the outside. The biggest hurdle for SMEs is basically the decision to deal with the topic.
Aren't big companies the real winners in developing these technologies?
The larger a process or the range of processes that exist in an organization, the greater the impact of automating those processes. Anyone who has processes that do not justify the implementation effort due to their minor importance or large variation can of course only benefit to a limited extent. That certainly applies to many smaller companies and craft businesses. In the work organization sector, however, smaller companies can also make sensible use of AI applications and benefit in the area of quality assurance, for example.
As a technology provider, do you see your duty to proactively demonstrate the advantages?
From our point of view, it is of central importance to take our customers with us and promote an understanding of the potential of deep learning. At senswork, we decided in spring 2020 that we wanted to do just that. In the future, we plan to position ourselves more strongly in terms of knowledge transfer on AI and to create a platform for this purpose. Specifically, these can be, for example, workshops for decision-makers (banks, funding agencies, IHK, HWK), discussion groups from teaching and research or classic train-the-trainer events for consulting companies. Courses for technically savvy and interested people who, over a weekend, will practically understand how the growth stage of the tomato plant on the windowsill can be assessed using AI are also conceivable.
Assuming there is a successful contact with a company: What is the further process, from the first ideas to the completion of the system?
In our domain, it has little influence on the general process whether an application is solved with conventional rule-based methods or a neural network. An application in industrial image processing is usually based on a specification, on the basis of which a decision is made regarding computing platform, camera, lighting and optical setup. The hardware plays a decisive role, because only if the data or the information contained therein is optimal, a high-performance evaluation with a high error detection rate is possible.
In addition, extensive visual material is important. If the image data are marked (tagged), a neural network is trained with them. As part of the training, the detection performance of the network is optimized. In the actual application, you work with the optimized network, which continuously receives and assesses individual input images during system operation. To qualify an application, it may be sufficient to use existing image material from the customer's production so that no special camera system has to be designed and set up in advance.
SMEs in particular are afraid of the financial risk, the loss of jobs or manipulated software. What are the reasons for your low commitment?
The points mentioned can be argued (which is often the case) with any type of automation. I don't think AI topics are special here. The biggest obstacle in my opinion is the lack of understanding of how neural networks work. There is a widespread belief that a complex computer system makes important everyday decisions that are beyond comprehension. The fear associated with this is unfounded in that almost every AI only knows what it has been trained to do by a human.
Can you show practical examples with regard to quality management in the implementation of your projects?
The use of deep learning is very effective in applications with objects that are subject to a large range of variation. This includes many naturally occurring materials and particles or plants and their growth. The assessment is usually very easy for a human specialist. Due to the high level of complexity, the implementation of such applications was only feasible in relatively simple configurations until recently. From a certain degree of complexity, a set of rules for the clear differentiation and classification of features can no longer be described, which means that a conventional algorithmic solution is a long way off. With the help of neural networks, there is now an approach to deal with this problem, since a description of rules is no longer necessary.
How is the relationship between large companies and medium-sized companies as customers?
We have a very balanced relationship. We develop systems for very large customers as well as for SMEs. Of course, this applies to image processing systems in general. For our customers, the solution to the task is in the foreground. For us, the application of neural networks is a major evolutionary step in improving analysis performance.