Not so long ago, I had a meeting with a customer at some outdoor location. They’re an innovative company developing an agricultural robotics application. Because they were dealing with a sensing problem, they wanted to demonstrate their machine to me in the field. Unfortunately, it had been heavily raining shortly before our meeting. Although my motto is ‘artificial intelligence doesn’t happen in the office behind your keyboard but in the field with your customer’ – I use this to motivate young talent to think further than Python solutions only – I wasn’t well prepared and forgot to bring boots beside my laptop. Using the right tools is always essential to achieve success.
The reason I was asked to visit the field was to assess whether AI technology could improve the sensing part of the robot. While the mechatronics worked just fine, the sensor system’s performance was lagging. Too many of the vegetables that needed to be detected were overlooked by the vision system, thereby decreasing the robot’s productivity and thus weakening the overall business case.
The field of AI has brought forth a new technology that’s becoming a powerful tool for engineers: deep learning. Next year, it’s going to celebrate its 10th anniversary. The basic ingredients of deep learning are big datasets, algorithms and huge compute power. The combination of these three things has already achieved amazing results. The latest deep-learning algorithms outperform traditional vision algorithms. Also, they’ve enabled the development of speech recognition assistants such as Siri, Alexa and Google Voice and they’ve developed better strategies than humans to win complex games such as Go.
Deep learning is a new versatile technology that solves problems that can’t be solved so easily by traditional engineering techniques – by which I mean ‘first principle’ approaches, whereby the engineer builds a model of a problem and tries to find a solution analytically. Such approaches aren’t always practical or possible.
I’m not a fan of ‘AI first’ and trying to solve any problem using AI. In my experience, artificial intelligence starts to be an effective tool when you’re dealing with a problem that has a lot of variation or where human intuition is the only way to solve it. Consider, for example, measuring the surface area of the leaves of a tomato plant. To solve this problem, a vision system needs to recognize leaves. Creating a model that describes what a leaf looks like is difficult, as leaves vary a lot in properties such as size, shape, color, texture, position, orientation, damage and thickness. Here, deep learning starts to shine. Deep-learning algorithms are very effective in processing large datasets with many different leaf variations and optimizing a vision algorithm to cope with these variations.
Variation and intuitive knowledge are two characteristics of agricultural systems. With the introduction of deep learning, a powerful technology is now available to develop advanced agritech applications, including smart plant monitoring, yield prediction and agricultural robotics. Agriculture is an ideal application field for deep learning to show its strength the coming years.
Looking back at my meeting, it was a perfect example of two seemingly different worlds coming together: muddy agritech and high-tech artificial intelligence.