When artificial intelligence goes deep

Publish Date

22 MAY 2021


Kai Beckmann


Today, artificial intelligence (AI) and the deep learning method enable surprising and spectacular applications, such as image and voice recognition.

At EMD Electronics, we are not only laying the foundations for the development and production of corresponding hardware with our materials and solutions in semiconductor technology; we are also applying deep learning ourselves and researching the principles.

Coming closer to human intelligence

Humans outclass machines in many respects. We not only experience emotions but are also one step ahead when it comes to intelligence. One common example of this is the ability of children to recognize and classify new objects in their environments very quickly. With only a small number of short interactions, children learn, for example, what a cat is – even though every cat is different, and they are seen from different perspectives or under different lighting conditions. Our brains are clearly able to abstract information from a few observations and generate a useful model of our environment without any special instructions. By contrast, informatics experts have been trying for a long time to teach computers these kinds of models through explicit programing – with limited success.

However, a lot has happened in the last ten years in the area of artificial intelligence. Artificial neural networks with a large number of interlayers, known as deep neural networks, and the deep learning method that is realized within them make this possible. With deep learning in particular, the computer largely learns the model for its “environment” by itself through statistical analysis of large quantities of training data rather than through small-scale pre-defined sets of rules. In this way, machines can recognize things, make predictions and take decisions – thus coming closer to human intelligence.

Deep learning has now enabled a range of practical applications and triggered a wave of enthusiasm for AI worldwide. Already today, we talk to chatbots and voice assistants that learn from every question we ask them and improve themselves. The figures also confirm that deep learning in practical application is no longer a distant prospect: According to a study by the market research and consulting firm Emergen Research, the global market for deep learning systems is to increase to €93 billion by 2028. Particular growth drivers are the adoption of cloud-based technology and usage of deep learning systems in big data analytics.

Deep learning at EMD Electronics

AI and deep learning require vast quantities of data and high computing power. Our Electronics business sector delivers the basis for this kind of powerful technology with semiconductor materials. Without them, powerful computer chips would be unimaginable. Examples of this are our contribution to 3D-NAND technology or our development work on neuromorphic computers. Nevertheless, deep learning still faces many hurdles. In particular, there often simply aren’t enough structured data for the algorithm to learn from. For this reason, deep learning and AI are currently used predominantly in scenarios that can be scaled en masse and for which large volumes of training data are available. For example, the automotive market and autonomous driving are particularly well-suited for this with millions of cars as data suppliers.

In the future, it could therefore be possible for an oncologist to upload a tissue image to the cloud and obtain a diagnosis within minutes. At Electronics, we are testing the application of deep learning in quality assurance. For example, we scan photos of samples of our liquid crystals, which are used for LCD displays. As a result, the AI is trained to recognize defective batches. These defective batches can then be removed quickly and effectively from the production process, which enables us to ensure even higher quality of our liquid crystals. 

Deep learning is also a topic when it comes to basic research Merck KGaA, Darmstadt, Germany: We are currently setting up research groups that address the issue of how biological brains and machines can abstract information – see the first example with the cat. By combining neuroscience, informatics and mathematics, we want to research how these capabilities occur and how they can be translated to AI.

I believe that AI and deep learning have enormous potential. It will be a long road to travel, though. At EMD Electronics, we intend to follow this path intensively.

By {{ authorText(item) }}

{{ item.publicationDate }}