|
|
|
“Incoherent Multiple Imaging for Parallel Optical Interconnection: Applications in Adaptive Neural Computing”
by Douglas J. Wiley
December 1992
Optical technology has shown commercial success only for long-distance applications and optical disk storage to date, although recent interest has centered on the development of optical technology for shorter-range interconnections and switching. Optical computing and signal processing have traditionally lacked high-quality spatial light modulator devices, but practical applications may be within reach with the newest optoelectronic technologies. Optical interconnection for computer backplane communications represents one short-term goal, but the paradigm of the programmable serial computer may be supplemented by alternative computing technologies such as parallel neural networks. These networks are `programmed' by a training session according to `learning rules'.
This work investigates the potential implementation of neural networks using free-space optical interconnection. A refractive optical experimental system was developed that implemented different `learning rules' specifying changes to the optical interconnections.
Experimental system components included video equipment (liquid crystal television, miniature CRT monitor) to present input signals to the optics. A lenslet array performed simultaneous, multiple imaging of an array of input signals encoded as an array of light intensities (pixels). An analog weighting was optically applied to each input pixel, and a video camera and frame grabber card in a personal computer collected the output, applying a pointwise nonlinearity. The system implemented a completely-interconnected analog-signal, analog-weight neural network.
The analog-signal experiments included neural networks implementing supervised learning algorithms, where an expected output is compared to the actual output (perceptron). Unsupervised learning networks (competitive learning), and two types of (non-learning) maximum-finding networks were also implemented. Optical hardware requirements and considerations of optoelectronic implementation are addressed.
An extensive characterization of system performance was accompanied by efforts to modify the neural network algorithms to be more tolerant of the inaccuracies of the optical system. Techniques of compensating for error were also explored using simple software algorithms, hardware, and signal encoding choices. The techniques of quantitative performance evaluation developed are applicable to a wide class of optical interconnection and neural network systems.
The experimental system successfully demonstrated learning in a hybrid optical/electronic system, and the experience of building a working system taught valuable lessons about the cumulative effect of many small inaccuracies (particularly crosstalk).