SPERT: a neuro-microprocessor

High-speed special-purpose digital systems for artificial neural network (ANN) calculations have been proposed by a number of researchers [Hammerstrom, 1990, Ramacher et al., 1991]. In 1989, our group designed and built a programmable machine called the Ring Array Processor (RAP) for general backpropagation training of layered feed-forward networks [Morgan et al., 1992]. This used multiple TMS320C30 DSP chips [Tex, 1988] connected by a ring of programmable gate arrays. We are currently developing a single chip CMOS microprocessor, called the Synthetic PERceptron Testbed (SPERT), to make fast and flexible neurocomputing more cost-effective.