Optical neural network computing is of great interest in terms of massively parallel computing. In recent years, CCD cameras, optoelectronic smart pixels and spatial light modulators (SLMs) with the high spatial resolution are reported[1,2]. In some cases, however, the interface between 2-D inputs and parallel neural computing systems or between the computing systems and output devices is not parallel but serial. The bandwidth of the interface between the I/O systems and the main computing system is limited and therefore this limits the performance of the total system. Such a problem is sometimes called I/O bottleneck. An all-optical parallel neural computing system with highly parallel I/O capability has been reported[3,4]. The system of the holographic associative memory, however, has limited functions and performances, because of less flexibility of optical systems. An alternative approach is to employ functional optoelectronic systems for wide-bandwidth input data, which can compress the data for the neural computing system. In this paper, we present network system consisting of an electronic parallel interface or preprocessor is described, and a generic interface device using nonlinear organic material for such a system is finally proposed.

No abstract