Kobold: a neural coprocessor for backpropagation with online learning
Martin Bogdan, Heike Speckmann, Wolfgang Rosenstiel · 2002
In this paper we propose an architecture of a neural coprocessor for on-board learning standard backpropagation. The hardware implementation works as a neural coprocessor connected to a personal computer by a special asynchronous interface. The coprocessor consists of several equal submodules representing one column of the neural net. So the architecture allows to compose any size of neural net depending on the specific application and, additionally, recurrency is allowed. Kobold speeds up the performance in contrast to earlier hardware implementations because of its new specialized control and communication structure. The subprocessors communicate asynchronously and locally with their nearest neighbours and synchronously by a global bus. The operations are controlled by dataflow. This means, a neuron calculates its weighted sum as soon as all inputs are available.