Machine Learning for Embedded Systems

The general trend in the field of artificial intelligence in recent years has been toward ever larger and higher-performance models. In neural networks in particular the number of layers and weights is increasing explosively. This is directly related to the establishment of deep neural networks. But is bigger automatically always better? The size of neural networks demands ever larger, more energy-hungry and more expensive hardware. To conserve resources and the climate Fraunhofer IMS is developing machine learning methods that work directly on smart sensors. This includes both inference and actual learning.

To enable online training on very small embedded systems, such as an 8-bit microcontroller, new concepts are necessary: it is about distributed learning. A microcontroller alone is not capable of implementing the entire learning process completely independently. However, the individual embedded systems are able to master this task together through networking. The resources for machine learning are distributed along classes, layers and data. Fraunhofer IMS is researching on the automated realization of distributed learning methods for a wide variety of embedded systems. The AI software framework AIfES, which is also our development, is used too.

Another research focus of Fraunhofer IMS is the intelligent feature extraction for smart sensors. Deep neural networks already have an included feature detection, but this comes at the cost of size, so that they cannot be used on small embedded systems. By means of feature extraction specifically adapted to the problem at hand, the network size can be kept as small as possible while maintaining the highest possible accuracy. These concepts have already been implemented in LiDAR sensors and for the detection of atrial fibrillation in ECG signals.

Machine learning always requires data that represent the requirements in the use case as realistic as possible. However, sufficient or high-quality data do not exist for all applications. One solution can be the use of synthetic data for learning. In order to be able to use synthetic data that is as realistic as possible in the learning processes, Fraunhofer IMS uses hybrid models in the form of Physics Guided Neural Networks (PGNNs). This merges the worlds of simulation and machine learning.

Fraunhofer IMS does not only research in the area of supervised learning methods but also in the area of unsupervised learning methods. One application example is Nonintrusive Load Monitoring (NILM) (Link: NILM). The aim is to measure the device-specific energy consumption without influencing the measurement system itself. Possible applications include energy efficiency, condition monitoring and predictive maintenance.