Performance evaluation of neural network hardware using time-shared bus and integer representation architecture

Neural network hardware using time-shared bus and integer representation architecture has already been fabricated and reported from the design viewpoint. However, nothing related to performance evaluation of hardware has yet been presented. Computation-speed, scalability and learning accuracy of hardware are evaluated theoretically and experimentally using a Back Propagation (BP) algorithm. In addition, a mirror-weight assignment technique is proposed for high-speed computation in the BP. NETTalk, an English-pronunciation-reasoning task, has been chosen as the target application for the BP. In the experiment, recently-developed neuro-hardware based on the above architecture and its parallel programming language are used. An outline of the language is described along with BP programming. Mirror-weight assignment allows maximum speed at 55.0 MCUPS (Million Connections Updated Per Second) using 256 neurons in the hidden-layer (numbers of neurons in input- and output-layers are fixed at 203 and 26 respectively in NETTalk). In addition, if scalability is defined as a function of the number of neurons in the hidden-layer, the machine retains high scalability at 0.5 if such a maximum speed needs to be used. No degradation in learning accuracy occurs when experimental results computed using the neuro-hardware are compared with those obtained by floating-point representation architecture (workstation). The experiment indicates that the present integer representational design of the neuro-hardware is sufficient for NETTalk. Performance has been evaluated theoretically. evaluation purposes, it is assumed that most of the execution-time is taken up by bus cycles. On the basis of this assumption, an analytical model of computation-speed and scalability is proposed. Analytical predictions agreed well with experimental results.