AI in 2020 From Experimentation To Adoption

Artificial intelligence is the umbrella term for applications in which machines perform human-like intelligence. This includes machine learning or machine learning, natural language processing (NLP) and deep learning. The basic idea is to use machines to approach important functions of the human brain – learning, judgment and problem solving.

This type of learning enables, among other things, so-called natural language processing (NLP). This involves the processing of texts and natural human language, which is used, among other things, by the Amazon voice service Alexa. Deep learning is currently seen as the most promising method of machine learning, which uses very deep neural networks with several levels and a large volume of data.

In contrast to NLP, the algorithm for deep learning goes deeper: The machine recognizes structures, can evaluate them and improve itself in several forward and backward runs. The algorithm uses several node levels (neurons) in parallel to make informed decisions. For example, medicine with deep learning finds support in the early detection of cancer or heart disease and can examine children’s DNA profiles for gene markers that indicate type 1 diabetes. In research, deep learning is used, among other things, to evaluate thousands of cell profiles and their active genes, or particle showers that arise when proton beams collide in a particle accelerator.

Since this type of learning solves complex, non-linear problems, it is also used in self-driving vehicles, for example, to correctly interpret confusing traffic scenes: pedestrians, cyclists, weather, traffic signs or trees – the behavior of road users must be correct, taking all possible influencing factors into account recognized and predicted.

However, the AI ​​congress not only offered visionary flights of fancy, but also “grounded” things – use cases and very specific applications. After all, AI – not just in the form of chatbots or smart speakers like Alexa – has meanwhile found its way into our everyday lives. More and more industrial companies are relying on AI, as company representatives demonstrated with use cases and best practices at the Handelsblatt Congress.

The Ditzingen-based mechanical engineering company Trumpf tries to solve a well-known and rather annoying problem with laser punching with AI support. So far, it has always happened that the parts “cut out” with the laser get jammed in the workpiece (ie sheet metal). This led to collisions in the machine and to production downtimes. With the help of special sensors and an AI algorithm, Trumpf has now found a way to avoid these disturbances.

Trumpf also tracks machine noise as an indicator of potential malfunctions and failures, as Klaus Bauer, head of basic technical research at Trumpf, explained in his lecture at the Handelsblatt Congress. A special smartphone app has recently been launched. In a first step, the machine sends the noise spectrum recorded by the microphone to the Azure Cloud, where it is analyzed using an AI application. For example, vibrations can be identified that the ear cannot recognize at all. The analysis results finally reach the app from the cloud. The service technician can then isolate the problem and decide whether and which maintenance or repair measures need to be taken.

AI needs data that is not big enough to train the algorithms. And of course computing power. There is no shortage of the latter in the cloud age, but data is often lacking. According to Klaus Bauer, the situation could be better if more companies decide to transfer their (machine) data to the cloud. But the data quality also leaves something to be desired. According to Dr. Bernd Heinrich, CDO for Smart Mobility Solutions at Bosch, about 80 percent of the data collected is simply worthless, which is why the pool of usable data for training the AI ​​engines continues to shrink. An alternative could be training with simulated data . Trumpf-Manager Bauer has the digital twin as a data source- in other words, the virtual image of a machine or system. This “digital twin” can be used to simulate certain machine states. The data generated in this way can then be used to train the algorithms and neural networks.

A survey conducted by the VDI (Association of German Engineers) among 900 members provides similar findings . So Germany’s position in the global AI scene is not really bad, but also not really “great”. According to 80 percent of the respondents, top AI nations are the United States, followed by China, with the Middle Kingdom accelerating its AI campaign as part of its “China 2025” plan. Germany comes in third place with a vote of 30.4 percent.

Like Dr. Kurt Bettenhausen , chairman of the interdisciplinary VDI committee Digital Transformation, when presenting the survey results at the Hannover Messe 2018, said that the USA is a leader in basic research, but also in the use of AI to evaluate unstructured consumer data, due to stricter data protection regulations such as the new GDPR in Germany and Europe is only possible to a limited extent. The Siemens manager who has been working in the USA for years sees this as “opportunity and risk” at the same time. “

According to the VDI study, the prerequisites in China are also completely different than in this country. In China, one has to deal with centralized structures. The business plan is investing heavily in AI, with the national goal of becoming the world’s No. 1 in artificial intelligence by 2030. Bettenhausen: “China has set a breathtaking pace as part of the digital transformation and is skipping some of the developments that Europe has undertaken.”

HOW DO MACHINES LEARN?

Machine learning either takes place through training using a data set with already known outputs (monitored), or algorithms themselves have to recognize patterns in data (unsupervised). It is also possible to learn through reward and punishment (enhanced), in which the algorithm automatically recognizes whether the learning component uses the entire system (reward) or not (punishment). The data are either structured, for example in tabular form, or unstructured as text, images or speech – as with emails or social media posts. Machine learning can process all data, which is a great advantage. (Source: SAP)