Seminar

Abstract: 

[nur in Englisch verfügbar] 

The power of deep neural networks (DNN) has been harnessed for many concrete applications and the potential of DNNs is still not fully explored. An especially interesting field of application is the embedded world, where »smart« devices can quickly analyze sensor data and produce meaningful responses.  However, if more of the analysis is done on the device, the energy consumption rises.

Therefore, the analysis must be made as efficiently as possible to keep the energy consumption low. Designing an efficient DNN by hand is cumbersome, and often it is not clear which design choices lead to a good model. Neural Architecture Search can find a good model for a given problem by searching through many, automatically generated models and selecting the best fit. We introduce our Neural Architecture Search Engine (NASE), which is designed specifically to include hardware constraints into the search and to find DNNs in a Pareto optimal sense.