Laser Neural Network Acts as Computer

Kevin Robinson

EINDHOVEN, Netherlands -- An optical processing network that uses the effect of external feedback on the output of a laser diode to control processor responses could prove to be the harbinger of the elusive optical computer. The researchers seek to mimic, in simplified form, the parallel processing abilities of biological neurons.
E.C. Mos of Philips Research Laboratories and H. de Waardt of Eindhoven University of Technology, led development of the parallel processing by splitting the beam of a diode laser into its longitudinal modes. Each mode represents one neuron that can have a number of inputs but only one output.
The inputs to each neuron are weighted and summed. This summation is compared with a threshold. Depending on whether the sum exceeds the threshold, the neuron's binary output will be either a one or a zero, on or off.

Beam is split
In the device, the beam of a multiple quantum well laser from Philips Optoelectronics is split into several beams that are then passed through a series of lenses, polarizers and a liquid crystal display, where they are attenuated proportionally to a weighted sum of the inputs. These attenuated beams are reflected to the diode, where the optical feedback they create is enough to modify the longitudinal mode spectrum of the laser diode. The optical output power varies according to this feedback.
The attenuation mechanism is a passive matrix liquid crystal display. Each pixel can be set to one of 31 gray-scale levels. The output power was measured by attaching an optical multichannel analyzer to the system.
Two personal computers communicate with each other to "train" the system. The first computer adjusts the gray-scale display, and the second monitors the power output. For each trial, the second computer reads the output and sends the results to the first. The first adjusts the gray-scale according to a learning algorithm until the differences between the actual and desired outputs are below preset levels.
Using this method, the researchers successfully trained the system to perform several logical functions. They also plan to train the network to perform high-speed pattern recognition, but first they must expand the input device to accommodate more than four or five inputs.
The group's complete work was published in the September issue of Applied Optics..