Embedded machine learning: micro-intelligence for sensors

| |
1 Star2 Stars3 Stars4 Stars5 Stars

The mission of electronics is to simplify our lives in the future. Artificial intelligence (AI) will certainly play a major role in this evolution. A new development is designed to incorporate AI into miniature devices without the help of a cloud link and PC power.

Micro-controllers will be the domestic heroes of our lives. They are now being used in just about every electronic device, ranging from smartphones and cars to washing machines and toothbrushes. Sensors frequently churn out the data on which machine-learning algorithms can be used.

But current software solutions use Python as a programming language and are generally available only on PCs. This means that neural networks cannot be trained and run on embedded systems with micro-controllers.

With the help of AIfES (artificial intelligence for embedded systems), a research team at Fraunhofer IMS has now developed a machine-learning library programmed in C that can be run on micro-processors and supports other platforms like PC, Raspberry PI or Android. It comprises a fully configurable artificial neural network (ANN) with a feedforward structure that also creates deep networks for Deep Learning. As a result of the reduced source code, the ANN can be trained directly on the embedded system.

“Micro-intelligent” embedded machine learning

Of course, embedded systems are unable to handle massive Deep Learning models. But a problem-focused pre-processing strategy for data and intelligent feature extraction facilitate a form of “micro-intelligence” at the very least. The small ANNs then facilitate subsequent learning on the controller itself.

The research team has already developed a demonstrator that can recognize handwritten numbers on an inexpensive 8-bit micro-controller (Arduino Uno). Another one is equipped with an absolute orientation sensor and can recognize complex gestures made in the air. During the training process, different individuals write the numbers zero to nine several times. The neural network learns this training data and identifies it independently in the next step. The process works with nearly all numbers. As a result, wearables could be equipped with gesture control, and people could protect their privacy.

The application potential of AIfES is virtually unlimited. For instance, a wristband with integrated gesture recognition could be used to control indoor lighting or monitor exercises and activities in a rehabilitation or fitness area. In the process, people’s privacy would remain protected because neither a camera nor the cloud were being used.

 

 

 

 

Embedded Machine Learning (Image: Fraunhofer IMS)

The AIfES recognizes handwritten numbers after undergoing brief training. (Image: Fraunhofer IMs).