12 October 2020
DSM and Lightyear aim to design solar roofs for all types of electric vehicles
The first edition of the AI Summit Brainport in Eindhoven was all about (regional) cooperation. AI (Artificial Intelligence) systems are going to play an increasingly important role in industry. However, before you can further implement AI, you have to overcome many obstacles. The AI Summit is organized by EASI in cooperation with AI Innovation Center, Province of North Brabant and other partners of AI hub Brainport. During the conference, researchers, professionals and AI enthusiasts come together to share and increase their AI knowledge and skills.
"In the beginning, machines helped our arms and legs. Today, they help our brains." With those words, Carlo van de Weijer, director of the Eindhoven AI System Institute opened the AI Summit Brainport. He also has a warning. "AI is growing exponentially. When we were leaders in the glass tube industry, we were not standing still. We have to do the same now." So there is a rush to develop and implement AI.
"Right now, there are very few AI systems in healthcare and the impact on patient care is limited. We are very good at inventing software, but not at implementing it. That's because we only have small and limited data sets at our disposal. That is why I am so pleased that there is now a movement to make more and more data publicly available."
One of the main issues that arises in this regard is patient privacy. Data is needed to train an AI, but without the patient's consent, the data can't just be used. De Jong cites two solutions during her keynote that would protect patient privacy.
One would be "Federated Learning. In this, the data is not sent, but stays in the same place. Instead, a party downloads the AI model through a central system to then train it with the available data. Once finished, the model is sent back and another party can work with it. This continues until the AI model is trained enough. Another way AI can be trained without invasion of privacy is to use synthetic data. "It is not real data, but resembles the original data. In fact, the structure of the data remains the same."
During the conference, Breemen will talk about his experiences implementing AI in industrial applications. For example, he started a collaboration with VDL in 2020 to create a visual system for a robot that cuts off leaves from cucumber plants, but leaves the vegetable itself alone. Breemen then shows a picture of what such a greenhouse with cucumber plants looks like. The leaves and crops hang crisscrossed.
"I call this a visual jungle. This is so hard to do, this is where you have to use Deep Learning."
Albert van Breemen, CEO and CTA of AI engineering company VBTI
The project has been running for two years. Breemen is only now starting to see really nice results. "An 80 percent success rate is easily achievable, but this robot has to cut a thousand leaves an hour, eight hours a day for 180 days per season. In that time, the robot is allowed a maximum of five errors. That's 5 out of 1.5 million. To get there, you have to have more data to train the model. That's the tricky part."
"In the beginning, we thought AI software creation was a simple linear process. You train an AI system and then you have a model ready to give to a robot, but how wrong we were," Breemen says. "Nothing could be further from the truth. You have to keep training the system for every little change, such as what light is used."
According to him, creating AI systems is actually a continuous process. "Customers should therefore see AI as a service and not as a product. We need to keep collecting data to perfect the AI model and make it applicable in other situations."
There is one aspect that cannot be missing from that: You need a broad and good team with different types of engineers and AI specialists, working well together.