'Scailable' showcases powerful combination of Edge AI and LoRa with KPN Things
June 14, 2021
The LoRa network is perfect for sending small data packets efficiently. But what if you want to have large amounts of raw data analyzed by artificial intelligence? “That is also possible by combining LoRa with Edge AI,” says Maurits Kaptein, CEO of Scailable. His company develops technology that allows even simple IoT devices to run an AI model.
Within IoT applications, it is common to send sensor data to a central location in the cloud. There, the data can be processed, for example, by machine learning models. “This takes a lot of electricity, computing power and bandwidth. Moreover, for many applications it is not necessary at all,” explains Kaptein. “It's more efficient to pre-process the raw data on location, because that's what AI actually does. Then you only send usable information and you minimize data traffic.”
One condition is that the AI must work on all kinds of different (types of) devices. “I thought there was no good technical solution for this,” says Kaptein, who is also Professor Emeritus of Data Science at Tilburg University. “More and more companies are developing AI models to interpret data. Think of counting people, recognizing license plates or detecting vibrations in mechanical parts. But how are you going to use such a model efficiently? Scailable can help you with this.”
Lower-level programming language Scailable uses two tricks. Kaptein: “A set of rules is often specified in a high-level programming language to make programming easier. But the higher the language, the slower the calculations. We strip an AI model down to almost machine level and tear out everything that is not specifically needed for a task. This makes the calculations much faster, sometimes up to 10 times faster. This also consumes less energy, because the computer runs for a shorter amount of time.”
The second trick is the use of interoperable technology. “Factories are full of industrial routers and gateways that often contain an ARM processor. It is very different from an Intel processor in a laptop or a processor in a camera. Our software automatically converts AI models into an efficient lower-order programming language from Mozilla. This language ensures that we can run the same AI model on all kinds of different devices. There is still a small layer between the hardware and the AI model, but that is minimal.”
Scailable's applications Scailable's technology automates the entire process to run an AI model on an IoT device - a model that has been trained on a laptop or in the cloud. That results in cost savings. “But Edge AI also lends itself to applications where there is no stable internet connection or there is a shortage of bandwidth. In addition, it allows you to detect anomalies in data quickly. And there are also companies that simply do not want to transfer all their data to the cloud, for security reasons for example.”
Kaptein mentions a number of examples of applications. “Companies use our tool, for example, to count how many cars are parked somewhere, to see whether people are keeping enough distance and to blur faces. But we also have many industrial customers. For example, we work for a large paper manufacturer. This company continuously collects data about its factories and machines. Thanks to Scailable, they can intervene quickly if the quality of the paper deviates from the standard somewhere.”
Emotion recognition via LoRa Scailable also appears to be an excellent match with the LoRa network, says the CEO. “You're not going to send raw audio or video over LoRa, that's not what this network is for. But if you process the data on location, you can transfer the results super efficiently.” Kaptein demonstrated this in practice during KPN's 19th Startup Afternoon. He used Scailable to run a fairly complex AI model for emotion recognition on a simple ESP32 chip with a camera.
Next, the ESP32 was paired with an Arduino MKR WAN 1310. This device sent the output – whether the person is laughing or not – via LoRa to the KPN Things Platform. “It worked great,” Kaptein recalls. “With this demo, we wanted to show how easy it is to combine Edge AI via KPN Things with LoRa. The average technician can get this working in no time.” The CEO has written an extensive manual with which every developer can repeat the demo themselves.
Starting small with KPN Things Kaptein enthuses over the KPN Things Platform . “It is the easiest way to use the LoRa network. Technically, the platform is simple. Developers have access via the KPN Things Portal to extensive documentation, in which, for example, the registration of devices is clearly explained.”
The CEO of Scailable expects more companies to make use of the combination of Edge AI and LoRa in the coming years. “KPN Things is an ideal starting point for these kinds of projects. It's one of the building blocks you need, just like the hardware and the AI models.” He advises developers to start with functionality in an IoT project. “And then you look for the right building blocks for that. You don't have to reinvent the wheel. Many building blocks are already available plug-and-play.”
“AI is wrongly seen as a kind of magic,” Kaptein concludes. “We can make computers do really nice things, but they are still sets of rules. So don't be intimidated and think especially in the applications that you can make with AI.”
Emotion recognition on an ESP32 with an AI model converted by Scailable into an efficient programming language. The results are sent to the cloud via LoRa using KPN Things.