Abstract

This paper concerns the development of artificial neural network
applications in digital signal processors (DSPs) with
multiprocessing capabilities. For defining an efficient partition
of processing tasks within target DSP boards, a user-friend
development environment evaluates different parallelism approaches
for the network design according to specific device features. The
development environment supports two different levels of
experience (novice, expert) on both parallel system and neural
network designs, enabling novice users to access system resources
easily and shorten design cycle time. Preprocessing methods based
on topological mapping and principal component analysis are also
supported by the system, so that compact and efficient neural
network system designs can be implemented. ...