Nuremberg, Conference  /  April 09, 2024  -  April 11, 2024

Embedded World 2024

Fraunhofer-Gesellschaft's joint booth (Hall 4, Booth 422)

The »embedded world Exhibition and Conference« is the industry meeting place for leading experts from the embedded community and industry associations. It offers an insight into the entire world of embedded systems: from components and modules to operating systems, hardware and software design, M2M communication and services. 

A team from the division »High Performance Computing«, the department »Financial Mathematics« and the University of Kaiserslautern-Landau (RPTU) will be presenting our AI system »Neural Architecture Search for Embedded Applications« (NASE) service at the Fraunhofer-Gesellschaft's joint booth (Hall 4, Booth 422).

Neural Architecture Search: The Link between Application and Hardware

The amount of data produced worldwide is constantly increasing. This also increases the need for efficient data processing in order to reduce energy and infrastructure costs. Many industrial plants, for example, have sophisticated sensor technology to control and monitor processes, the data from which is stored in their own cloud systems. In the age of the Internet of Things and 5G, however, edge computing is becoming increasingly important. Data is processed decentrally, i.e. »at the edge of the network« – directly locally instead of in remote data centers. Processing is shifting to the edge areas, directly where the data originates. However, this poses major challenges: How do we create the connection between application and hardware that scales with demand?

One promising technology is machine learning – in particular deep neural networks (DNN). They are very flexible and process a very wide range of inputs, e.g. images, videos, text and speech. However, attempts to transfer off-the-shelf DNNs to an embedded system often fail because the models are too large unless they have been specifically developed for the limiting case. Current and promising techniques to automatically search for optimal DNNs are AutoML and Neural Architecture Search (NAS). NAS searches for a DNN structure so that the DNN fits the hardware and fulfills the application goals.

We at Fraunhofer ITWM are developing a hardware-based NAS. This type of search also takes into account the limitations imposed by the hardware platform and we find solutions that run optimally. Users only provide the data and need little to no knowledge of DNN. Since both the search and the training of the NAS are done automatically, this is a very scalable method to quickly get solutions for the hardware.