Thin-film neuron network with optical programming

Details Thin-film neuron network with optical programming Thin-film neuron network with optical programming

1992-02-28

1993-05-18

Nelms, David C.

Radiant energy

Photocells; circuits and apparatus

Photocell controlled circuit

395 25, H01J 3150

Patent

active

052123774

DESCRIPTION:

BRIEF SUMMARY
BACKGROUND OF THE INVENTION

The invention relates to a thin-film neuron network with optical programming.
Conventional von-Neumann computing machines operate in accordance with a mathematical rule which is predetermined with the aid of a program. The prerequisite for solving a problem with a von-Neumann computer is therefore that the approach to the solution is known in principle. This is because it is only for an approach known in principle that an algorithm can be set up which is converted to a program.
There are a number of problems for which it is not possible to specify a precise approach in principle to a solution. For example, when comparing a signature with a stored signature sample, a correspondence must be detected even if the two signatures differ slightly. A von-Neumann computer with conventional programming reaches the limits of its capability with such problems of recognizing incomplete patterns.
By contrast, problems of this type can be solved by using neuron networks, also called neurocomputing. Neuron networks function in accordance with the same principle as the human brain. A neuron network consists of a multiplicity of neurons which in each case have many inputs and one output. The neurons are connected to one another via links of different strength which are called synapses. In the neuron, the weighted sum of all inputs is then formed. From this weighted sum, an output signal is generated via an evaluator which is a component of the neuron. The weighting of the individual inputs occurs via the synapses, the links of different strengths. In the evaluator, the sum is evaluated in accordance with a predetermined function. Functions considered are, for example, step functions, threshold functions or the sigmoid function. Neuron networks are described, for example, in J. J. Hopfield, IEEE Circuits and Devices Magazine, September 1988, pages 3-10, and Design and Elektronik, issue 20 of Sep. 27, 1988, pages 94, 96, 100-102, 104, 109-117.
The information in a neuron network is stored in the strength of the synapses. The strength of the synapses is initially defined with the aid of problem examples having a known solution which run in the neuron network. During the operation of a neuron network, the synapses are adapted with the aid of solutions found for problems processed. This process corresponds to the storing of experience in the human brain. That is to say a neuron network "is continuously learning more".
For a neuron network to have approximately the capacity of the human brain, the corresponding number of neurons must be implemented. The surface of the human brain amounts to about half a square meter in area. This large area required for an efficient neuron network is the reason for the hardware implementation still being in its infancy even though a high-performance and mature technology can be seen, for example, in microelectronics. This is why neuron networks have previously been mainly simulated on computers.
A main problem in the hardware implementation consists of the representation of the synapses. In principle, programmable resistances which are of simple construction and absorb little power are needed for the synapses since they are needed in a very great number.
The hardware implementation of a neuron network in amorphous hydrogen-containing silicon (a-Si:H) appears to be promising since amorphous hydrogen-containing silicon offers a number of characteristics which are particularly advantageous for implementing neuron networks. Essentially, the maximum size of a layer area in a-Si:H technology only depends on the size of the reactor used for producing the a-Si:H layers. It is therefore quite conceivable to create half a square meter of area corresponding to the surface of the human brain. The components needed for a neuron network such as resistors, diodes, transistors are of simple construction in this technology and therefore only require a small number of process steps. This is essential for the yield in the production of large, high-performance networks. In addition, the

REFERENCES:
patent: 4809193 (1989-02-01), Jourjine
patent: 4985621 (1991-01-01), Aull
"Artificial Neural Networks", J. J. Hopfield, IEEE Circuits & Devices Magazine, Sep. 1988, pp. 3-10.
"Design and Elektronik", issue 20 of Sep. 27, 1988, pp. 94, 96, 100-102, 109-117.
"An Optically Programmed Neural Network", by C. D. Kornfeld et al., IEEE International Conference on Neural Networks, San Diego, Calif., Jul. 24-27, 1988, pp. 357-364.

Affiliated with

Also associated with

Rating

Thin-film neuron network with optical programming does not yet have a rating. At this time, there are no reviews or comments for this patent.

If you have personal experience with Thin-film neuron network with optical programming, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Thin-film neuron network with optical programming will most certainly appreciate the feedback.

Rate now

Comments { 0 }

Profile ID: LFUS-PAI-O-806165