CD Laboratory for Embedded Machine Learning

Embedded Systems wie FPGA Plattformen haben großes Potential für Machine Learning Aufgaben mit hoher Performance bei niedrigem Energieverbrauch.
Rechenintensive Algorithmen wie Deep Neural Networks können Objekte erkennen wie hier im Foto von Laborleiter A. Jantsch mit seiner Forschungsgruppe.

Here, learning systems in the field of image and video recognition are being researched. The focus is on portable applications for outdoor use without access to large storage capacities and energy sources, e.g. automatic traffic light control systems.

 

Lifelong learning systems, e.g. systems that can adapt to changing scenes in real time and without human supervision, represent a lively area of research in the field of computer vision. Such systems are not trained in advance with large amounts of data on possible situations, but learn throughout their lives based on a few example situations in their actual working environment.

At present, however, such systems are mainly used for object recognition and are based on hardware that is not subject to any restrictions in terms of energy supply and storage space.

 

However, applications such as autonomous driving, augmented reality data glasses or automatic traffic monitoring systems require systems that adapt to a changing environment with little human intervention, that function under extreme weather conditions such as rain, snow and dirt, are portable and can manage with limited memory capacities and limited power supply. An example of such a complex application is the monitoring and automatic switching of traffic lights at a pedestrian crossing, where pedestrians need to be recognised as such and their presumed walking direction determined, regardless of daily and seasonal changes and any temporary roadworks. Embedded Field Programmable Gate Arrays (FPGAs), Graphical Processing Units (GPUs), etc. are used for such applications.

 

The CD Laboratory conducts research at the highest level and develops design methods and network architectures that are

(1) Achieve maximum accuracy for a given energy budget,

(2) show the lowest energy consumption for a given target accuracy, and

(3) have the capability for continuous in-device learning.

Durch hoch-optimierte Algorithmen, maßgeschneidert für ein Embedded System, kann die Leistung bei gleichbleibender Genauigkeit dramatisch erhöht werden. Quelle: Wess, M., Dinakarrao, S. M. P., & Jantsch, A. (2018). Weighted Quantization-Regularization in DNNs for Weight Memory Minimization Toward HW Implementation. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 37(11), 2929-2939.

Christian Doppler Forschungsgesellschaft

Boltzmanngasse 20/1/3 | 1090 Wien | Tel: +43 1 5042205 | Fax: +43 1 5042205-20 | office@cdg.ac.at

© 2020 Christian Doppler Forschungsgesellschaft