Boulder, Colo. — SINCE they were first invented, the characteristics of three-dimensional pictures called holograms have made them a popular metaphor for human memory. Now, by incorporating holograms into a simple optical system that produces behavior reminiscent of recollection and daydreaming, Prof. Dana Anderson of the University of Colorado, Boulder, is helping make the connection more than purely metaphorical.
For the last three years, the young investigator has been exploring an unusual byway in the burgeoning field of neural networks. Neural networks are a new type of circuitry that models certain aspects of human memory more closely than does the digital logic typical of computers. Instead of the mainstream approach of building such networks out of electronics, Dr. Anderson is using mirrors, lasers, and holographic materials.
``The idea of bringing optics and neural networks together is potentially a very powerful one,'' says Prof. John Hopfield of the California Institute of Technology, the scientist who laid the groundwork for the recent interest in neural networks. ``A number of people are trying to duplicate electronic circuits with optical techniques, but that doesn't seem too promising. Anderson - and a few others - are trying to integrate optics and neural nets in a more fundamental way.''
Anderson's research began when he noticed there were a number of striking similarities between the mathematics of neural networks and certain optical systems.
The core of Anderson's neural network is an optical resonator. It consists of mirrors and lenses that create a closed path for laser light. As a light beam travels around this path, it passes through a cube of special holographic material. It also travels through a second cube, where the beam is strengthened by energy from a second laser.
Like other neural networks, his optical system exhibits the rudiments of associative memory. Human memory is associative: A person can recall a memory when prompted by partial, distorted, or related information. That is not the case with computer memory, which stores information at specific locations. The correct address or an exact match is usually required to retrieve stored data.
Holograms store information differently. If a hologram is broken, each of the pieces can be used to reconstruct the entire image. The smaller the piece, the less distinct the image. The distributed nature of holographic image storage is similar to the non-localized way that memories appear to be stored in the brain.
Anderson utilizes another remarkable aspect of holography to build his associative network. Two objects are illuminated with laser light so that the reflected beams cross in the holographic media and are recorded. Then, when light from only one of the objects impinges on the hologram, the image of both objects is reconstructed. Thus, the two objects are associated in the holographic memory.
The researcher has shown that such associated images can be recalled by shining a laser beam carrying a partial or distorted image into the optical resonator. Most of the time, but not always, the images associated with the input image brighten while other images stored in the hologram are extinguished. The reason for this is an effect called ``mode competition.''
``The easiest way to understand this is by means of a biological example,'' Anderson says. Take a resonator with two stored images. As the laser light propagates around the ring, the two images compete for energy in a fashion similar to two species of predators that compete for the same prey. When one predator gains an advantage over the other, it takes food away from its competitor. Similarly, when one stored image brightens, the others weaken. Thus, injecting light into the resonator that strengthens even a portion of one image more than the others gives it an advantage that allows it to extinguish its competitors.
``The important point here is that if part of an image wins, the entire image wins. So, if the input image looked like several stored images, the memory recalls only the stored image with the greatest similarity,'' the scientist explains.
But what happens when no images are input? The result is unique to the optical resonator. It cycles through its stored images in a fashion reminiscent to daydreaming. ``No other neural network does anything without input,'' says Anderson.
The scientist freely admits that, thus far, his experiments have been very primitive. ``So far we have only been able to demonstrate this using boring spots as the images,'' he says. Currently, technical limitations keep the system from working with more complicated images. On the other hand, he has completed theoretical studies that suggest the system should also display behaviors similar to forgetting and obsession.
``It's remarkable that such a simple, optical circuit has so many similarities to human memory,'' Anderson observes. In the future he intends to begin building more complicated systems by combining optical and electronic circuitry.