Artificial Intelligence (AI) is getting more and more important. This is due to the increasing availability and volume of data and the reduced costs for the required computing power. The first relevant technology drivers and financial sponsors were the marketing industry (personalization in the digital world) and the defense sector (autonomous systems for military use).
An essential aspect of AI in an industrial context is learning. This means to deduce the solution to new or hitherto unknown problems based on known or learned solution strategies (learned strategies). In the following discussion, we will ignore the idea of completely emulating human intelligence in all its aspects (the domain of the so-called Artificial General Intelligence or AGI).
The AI subdomain concerned with learning is called Machine Learning (ML). ML in turn comprises subdomains like artificial neural networks (ANNs) with yet further subdomains like Deep Learning (DL). There is a great variation not only in the different ML structures but also in the objectives and the training, whether supervised (mappings provided) or unsupervised or reinforcement learning.
Machine Learning makes it possible to detect and classify anomalies, picture contents, speech, tests, technical relationship and the like in a reliable, robust and efficient manner or to utilize (unknown) relationships. In the media domain, this is also described as pattern recognition. In the simplest possible case, a neural network creates a mapping (a model) between an input (e.g. the picture of an apple) and an output (object: apple – probability 92%; object: pear – probability 3% etc.). Ideally, this mapping also works when the apple has a different color, is missing a bite, looks blurred or is shown upside down. For many specific, limited problems, the technology is superior to human perception.
The ML models first have to be trained for their respective tasks. This training is usually based on large prepared and selected data volumes and carried out on specialized servers providing the required concurrency by means of GPUs or FPGAs. When the system has been trained, it can be transferred to the target application. The big challenge is to make the pre-trained algorithms work on systems with fewer resources without incurring dramatic performance losses. This is always the challenge with embedded systems and hence with autonomous control systems, and it requires technological competency.
To evaluate the effectiveness and benefit of Machine Learning, it is essential to understand that ML tools have to be viewed as tools that can be used for specific problems. The rest of the system or the process remains untouched by AI. This leads to so-called hybrid systems as combinations of existing system components and ML tools. ML tools must therefore never be considered separately but always in the context of the overall system.
Machine Learning algorithms lead to performance increases in many fields. This increase is no longer the result of the hitherto steady increase in the number of transistors (the origin of Moore’s Law) but caused by the way in which the (probabilistic) systems detect, decide and act in the background.
In particular with ML algorithms like ANNs, it is very difficult to derive specific outputs from the input information (i.e., forecasting behavior) or to gain the specific decisive features based on an output. This is due to their architecture and complexity. This behavior is generally described as a black box. For more transparency, which is particularly important for the acceptance of autonomous driving, (empirical) research areas like Explainable AI were created.
In the upcoming years, there will be many performance leaps through technological change in Machine Learning, e.g. by memristor arrays, spiking neural networks, distributed training, optical networks or Turing-incomplete architectures. This will make it possible to solve ever more complex problems on site or on the machine, respectively. We are looking forward to the extensive positive consequences for our core business competency!