What if a system could use machine learning to train models and run them on the same microcontroller? It’s, in essence, the groundbreaking accomplishment of NanoEdge AI from Cartesiam, a French company and a member of the Machine Learning ST Partner Program that’s redefining what we know about artificial intelligence. As a software company, Cartesiam listens to their customers describe what they want to analyze (i.e., light, sound, electrical properties, etc.) and the available hardware (i.e., sensors, memory, MCU, and more). The company then delivers a library that enables the future application to take advantage of machine learning at the edge. The process itself is straightforward because the company has years of research and experience. Let’s, therefore, look at what our collaborator brings to the ST Partner Program, to IoT, and machine learning applications with a glimpse at how they used our latest STM32G4 32-bit MCU Series.
What is Cartesiam Bringing to the ST Partner Program? NanoEdge AI and Unsupervised Learning
The presence of Cartesiam in the ST Partner Program is more crucial than ever because it complements our initiatives. At the beginning of the year, ST launched STM32Cube.AI to enable developers to convert neural networks into optimized code for STM32 easily. Our tools target applications that rely on predetermined events. Developers train a neural network by collecting data before processing it in a neural network training framework on a PC to recognize specific activities, such as walking, running, or swimming. This supervised learning phase outputs a trained neural network that developers can then send to STM32Cube.AI to convert it into a code that will enable our MCUs to recognize these activities (i.e., inference phase). STM32Cube.AI is still today the only solution of its kind for STM32 MCUs.
On the other hand, Cartesiam offers a unique solution for customers that can’t know what to expect and can’t, therefore, run supervised learning sessions in advance. NanoEdge AI is original because it runs the learning phase on the microcontroller itself and without requiring complex frameworks on a PC. Engineers turn to this solution when they can’t create neatly pre-trained models for specific situations, but still desire to use machine learning to come up with smart solutions, such as predictive maintenance, despite the constraints in resources inherent to embedded systems. They can run the training phase on the MCU to learn the normal behavior of a device in its final environment instead of a lab, then run inferences on the same MCU to detect and report behavioral anomalies.
What is Cartesiam Bringing to the ST Partner Program? NanoEdge AI and STM32Cube.AI Hand in Hand
Cartesiam’s solution complements ours because the company relies on a completely different mathematical paradigm. Up until now, the industry worked under the assumption that training powerful machine learning models was only efficient on PCs running TensorFlow or Caffe, to name only two. Today, NanoEdge AI breaks this a priori thanks to a framework that uses new mathematical models that take into account the resources available on a microcontroller. While ST changed the industry by bringing trained models to STM32 MCUs, Cartesiam is a crucial ST partner because it now brings our microcontrollers to machine learning to open them to a whole new range of applications thanks to its ability to run unsupervised learning and inferences on one MCU. The company’s solution is also the fruit of years of research that ended up on our SensorTile module during Embedded World 2019.
What is Cartesiam Bringing to the Internet of Things?
NanoEdge AI and STM32 to All
To learn more about NanoEdge AI and its implementation on an STM32 platform, we sat down with Joel Rubino, CEO of Cartesiam, and Francois de Rochebouet, the company’s CTO. As they explained:
“We started experimenting with the SensorTile before we even became a member of the ST Partner Program because one of our customers had a similar design, and the module quickly became an excellent development platform. It offered us everything we required to create smart and connected prototypes without the need to come up with an original design.”
Cartesiam is thus a great example of how a company can innovate and change the industry when it doesn’t need to spend inordinate amounts of time and resources to create new hardware systems. Their demo at Embedded World was impressive as they showed how their machine learning libraries could use our SensorTile module to learn the behavior of a BLDC motor through vibration analysis, then detect and report an anomaly thanks to the embedded STM32L4 ultra-low-power microcontroller.
Behind the scenes, there was also another aspect of the demonstration that the public didn’t get to see, but that is a crucial advantage of Cartesiam’s solution: its ease of use. It took only four hours for Francois to implement the demonstration and developers can expect to integrate the Cartesiam library in their application relatively quickly. Developers get an example code from the French company, which dramatically lowers the learning curve, and will guide them as they call the learning function in a loop to start training the system before running a detection routine that rests on the model they just created. NanoEdge AI thus removes a lot of the complexity inherent to machine learning to make it accessible to more customers and more applications.
What is Cartesiam Bringing to Machine Learning? NanoEdge AI and STM32 Everywhere
NanoEdge AI is also an attractive solution because it is highly flexible. The solution can take data from all sorts of sensors, making it an excellent fit for a lot of industries. It’s the reason why we gave Cartesiam an early access to our new STM32G4 microcontroller to see what they would be able to do with all the optimizations we brought to the analog and digital peripherals, and they didn’t disappoint. As Joel and Francois told us:
“During our time with the STM32G4 Mainstream Microcontroller, we ported our algorithms and frameworks onto the new architecture to perform current monitoring operations and create a predictive maintenance solution relying on our machine learning solutions.”
Being able to drive the motor and run the AI for the predictive maintenance system with the same MCU and at the same time is a lot more cost-effective, robust, and compact. ST is now working with Cartesiam to ensure that upcoming packs and development boards will run demo applications that use NanoEdge AI libraries to better bring their complementary solutions to our community, thus positioning the STM32 platform at the center of the machine learning revolution.
This example shows that NanoEdge AI’s is malleable because of its ability to leverage the “machine” in “machine learning.” Instead of mimicking human behavior to “see” a problem by using a camera or “hear” it with a microphone, Cartesiam used current sensing tools and analog peripherals to create a model that’s far more efficient. By offering a flexible solution that can adapt to a tremendous number of situations, the company can meet a vast array of applications, and we are proud to collaborate with them to ensure that STM32 will be a driving force in this new journey.