Markt&Technik: Herr Dr. Wenzel, what do the terms deep learning, machine learning and artificial intelligence mean for you personally in the context of assembly inspection, and where do you see the most important differences?
Dr. Thomas Wenzel: Artificial Intelligence (AI) is a generic term for all systems that are able to imitate human learning behavior. Machine Learning (ML) is a sub-area of AI and describes mathematical procedures that allow a machine to generate knowledge from data and information as independently as possible. Deep Learning (DL) is an ML procedure based on neural networks inspired by the functioning of the human brain, in which neurons learn and store knowledge in the form of connections. This occurs in neural networks with the very large amounts of data that are needed to construct a knowledge base. It is therefore not possible to point to a distinguishing feature. On the contrary, one is the specialization of the other. Artificial Intelligence is becoming increasingly important in all technical application areas. In particular, machine learning using deep learning with neural networks is gaining ground in pattern recognition applications. They are increasingly replacing classical image and signal processing methods, which are based on deterministic operations such as linear and non-linear convolution. In particular, if the task were to identify deviations that were not specified a priori as deviations, the ML procedures would come into their own. What appears as a pattern yet does not represent a deviation per se often goes lost unnoticed in classic approaches.
What opportunities do these technologies open up in assembly inspection?
We are seeing two major application areas for AI processes not only in assembly inspection, but also in the inspection of cast parts for example. On the one hand, there is automatic defect detection - in the area of X-ray based assembly inspection called AXI, Automated X-ray-Inspection. Increasingly simplified learning procedures enable fast training and quick adaptation to changing tasks and therefore ever more robust solutions. The other application area is not about the inspection of a single printed circuit board, but takes place one level higher, namely at the process level. This is about identifying changes in the process: What changes from sub-assembly to sub-assembly?
What kind of patterns can be identified in these changes?
These questions are well suited to deep learning techniques. Process changes especially trends that indicate any deterioration in quality, can be detected at an early stage. Proactive countermeasures can then be taken before critical deviations or even rejects occur. In this way the inspection system changes and becomes a smart sensor that provides important information for optimum process control.
Where are Deep or Machine-Learning and AI currently being used?
We already see AI - and DL in particular - in the area of automatic defect detection for all kinds of inspection tasks. Since the requirements placed on inspection are subject to constant change and the number of inspection tasks grows dramatically every day, DL methods have a distinct advantage over classical image processing methods due to their relatively straightforward adaptability. The areas with which most consumers are familiar are voice and face recognition, which many mobile phone users use on a daily basis. As Internet users, we have to contend with the results of a behavioral analysis each time we are presented with a range of "ideal" products. All this is based on the use of AI.
Which are the challenges and where are the limits?
From our point of view, the main challenge is in the interpretability of the results obtained from an AI system. For example, the exact reasons for any decision made by a neural network in assembly inspection are not always comprehensible. A neural net is usually a black box that does not like to be scrutinized. A poor decision can often only be corrected by improved training. The kind of in-depth analysis which is possible with classical classification methods does not seem to be within AI’s reach at present. The limits of Artificial Intelligence systems lie in the learning data: Information that is not contained within this data cannot be reliably recognized or generated in the application of the procedure.
What risks lie in the development towards Deep or Machine Learning and AI?
AI is currently experiencing a higher degree of trust. The performance of Deep Neural Nets (DNN) for example, is impressive. At the same time, however, concerns are rising. These range from far-reaching automation of R&D and manufacturing and the associated loss of jobs to possible complete monitoring through face and voice recognition to automated decision-making in military applications. At the end of the day it’s an ethical question: Which decisions do we, as a society, want to leave to Artificial Intelligence? Technically, there seem to be no limits at the moment.
What place do Deep-learning, Machine-learning and Artificial Intelligence have on your technology roadmap?
As a company that has played a pioneering role in automatic image processing systems for the X-ray field for many years, we have awarded ML a firm place on our roadmap. We will follow both of the approaches described above with ML, i.e. automatic defect detection and the evaluation of the production process. We will use ML to provide valuable information for process optimization.
What does the future hold for these technologies - in general and in assembly inspection in particular?
Artificial intelligence will continue to permeate our lives. Be this in a Smart Home, at work or on the road. There is no limit to the number of applications, and every day new ideas about how AI can make things simpler and faster crop up. Where this will all stop will not be determined by technology, but by society. The world of inspection is still in its infancy with regard to AI. The use in individual systems can only make inspection more robust and adaptable to new tasks. The total networking of all systems and thus of the relevant sensors will make possible a self-regulating process that operates almost error-free.
This interview was conducted by Nicole Wörner, Markt&Technik