Bacteriorhodopsin synapses use thermal gatingResearchers at the University of Nevada (Las Vegas, NV) have demonstrated that bacteriorhodopsin (BR) spatial light modulators (SLMs) can be used to interconnect adaptive neural networks. Bacteriorhodopsin devices are commonly promoted as optically ad dressed SLMs, but they can also be modulated by varying their temperature. Not only can such modulators be written on with analog patterns of light, the sensitivity of the material to these patterns can
Bacteriorhodopsin synapses use thermal gatingResearchers at the University of Nevada (Las Vegas, NV) have demonstrated that bacteriorhodopsin (BR) spatial light modulators (SLMs) can be used to interconnect adaptive neural networks. Bacteriorhodopsin devices are commonly promoted as optically ad dressed SLMs, but they can also be modulated by varying their temperature. Not only can such modulators be written on with analog patterns of light, the sensitivity of the material to these patterns can be varied. Because of this, BR-SLMs can be used as flexible synapses in neural networks that can learn, selectively, while being used--not just in the training phase. How much a particular synapse is meant to learn can be controlled by neural feedback, which allows flexibility while preventing information from being erased or corrupted with useless data.
Neural networks are systems that both store and process information. Often, the learning and performance of particular tasks are kept separate. First the network is trained using a representative data set with known outcomes, and the system incorporates each new piece of information into what it already knows about the problem. Once this process is finished, the network can be used. The result is an unchanging, but sophisticated, information filter, which has the advantage of being relatively easy to implement optically.
Holographic neural networks, for instance, are often trained on images and then fixed (either chemically or otherwise) before being used. This prevents the light used to send information through the network from destroying the data that are already stored.
David Shelton, at the University of Nevada department of physics, has found a way to allow neural networks to be both used and trained simultaneously. Bacteriorhodopsin is a photocyclic material that starts off as yellow. At room temperature, absorbing yellow light causes it, after progressing through several stages, to turn blue. Blue light then turns it yellow again. The material can, therefore, be used to modulate one wavelength of light based on its absorption of another--it is an optically addressed SLM.
Below about 100 K, however, bacteriorhodopsin`s optical properties are essentially frozen--its sensitivity to light drops with ambient temperature. Shelton determined that this gave the material an important additional degree of freedom, making it particularly suitable to encode synapses: the set of interconnection weights that determine which incoming signals are enhanced or amplified and which are diminished before they are summed by a particular neuron.1
Together the proposed neuron and synapse consist of a BR film connected to a heater, photodetector, integrated circuit, and vertical-cavity surface-emitting laser (VCSEL). The incoming signals represent the outputs of other neurons in the network (see figure, right). The BR-SLM modulates the intensity of the incoming beams of light and is modulated by them, depending on previous inputs and the device temperature. The accumulated light energy, summed using the photodiode, determines the electronic activation--this determines the behavior of the electronic chip. The chip, in turn, causes the VCSEL to turn on (or not), passing information onto other neurons. It also determines the temperature of the BR film. Shelton says that, because of the high-resolution of BR-SLMs, real devices should be able to encode u¥to 108 interconnection weights per square centimeter.
Simulations of the system have, so far, worked well. When presented with random data, weights for inputs with data that were correlated increased, but fell for uncorrelated data.
SUNNY BAINS is a scientist and journalist based near San Francisco, CA; www.sunnybains.com.
1. D. P. Shelton, Opt. Lett. 22(22) (Nov. 15, 1997).