1995-04-24
1998-01-06
Downs, Robert W.
395 24, G06F 1518, G06G 760
Patent
active
057064032
DESCRIPTION:
BRIEF SUMMARY
TECHNICAL FIELD
The present invention relates to a semiconductor neural device, and in particular, provides a semiconductor integrated circuit neural network possessing a self-teaching function.
BACKGROUND ART
Neural networks are the electronic circuits which model the nerve circuit nets of the brain; such neural networks show great promise in next-generation information processing. In order to create high-level information processing closer to that of human beings, attempts have been made to realize neural networks on ultra LSI chips.
However, there are still a variety of problems in realizing neuron computers on LSI chips using current semiconductor LSI technology, and the current state of affairs is such that the goal of application has essentially not been achieved.
The technological problems involved in realizing such computers in LSI chips are stated below.
Human brains have extremely complicated structures and an extremely high degree of function; however, the basic composition thereof is extremely simple. That is to say, the brain is comprised of nerve cells having a calculating function, termed neurons, and nerve fibers, which serve to transmit the results of such calculations to other neurons, that is to say, which serve as wiring.
A simplified model of the structure of the fundamental unit of the brain is shown in FIG. 29. References 901a, 901b, and 901c indicate neurons, while references 902a, 902b and 902c indicate nerve fibers. References 903a, 903b and 903c are termed synapse junctions; for example, such a junction applies a weight wa to the signal sent along nerve fiber 902a, and inputs this into neuron 901a. Neuron 901a determines the linear sum of the signal strengths inputted thereinto, and if the total value thereof exceeds a threshold value, the neuron becomes active, and outputs a signal to nerve fiber 902b. If the total value is less than or equal to the threshold value, the neuron does not output a signal. When the total value reaches or exceeds the threshold value, and the neuron outputs a signal, this neuron is said to "fire".
In an actual brain, these calculations, transmission of signals, application of weighting, and the like are all conducted by means of electrochemical phenomena, and the signals are transmitted and processed as electrical signals. The process of learning in human beings can be understood as the process of altering the weighting in the synapse junctions. That is to say, with respect to a variety of combinations of inputted signals, the weighting is slowly corrected so as to obtain a correct output, and the system finally settles at the optimal values. That is to say, human intellect is engraved in the brain as synapse weighting.
A large number of neurons are connected to one another via synapses to form one layer. It is known that six layers exist, one on top of the other, in the human brain. The realization of this structure and function as an LSI system using semiconductor devices is the most important problem in the realization of a neuron computer.
FIG. 30 (a) serves to explain the function of one nerve cell, that is to say, one neuron; this was proposed as a mathematical model by McCullock and Pitts in 1943 (Bull: Math. Biophys. Vol. 5, p. 115 (1943)). Research into the realization of this model using semiconductor circuits and into the construction of a neuron computer is being widely conducted even now. References V.sub.1, V.sub.2, V.sub.3 . . . V.sub.n indicate a number n of input signals which are defined, for example, as voltage sizes; these correspond to the signals which are transmitted to other neurons. References w.sub.1, w.sub.2, w.sub.3, . . . , w.sub.n are coefficients expressing the strength of junctions between neurons; these are termed the synapse junctions in biology. The function of the neuron is to output a value of "1" when a value Z resulting from the linear addition of all inputs V.sub.i with the weights w.sub.i (i=1-n) applied thereto exceeds a predetermined threshold value V.sub.TH *, and to output a value of "0" when the value
REFERENCES:
patent: 5052043 (1991-09-01), Gaborski
patent: 5129038 (1992-07-01), Kohda et al.
patent: 5239618 (1993-08-01), Yamaguchi et al.
patent: 5258657 (1993-11-01), Shibata et al.
patent: 5347613 (1994-09-01), Chung et al.
patent: 5465308 (1995-11-01), Hutcheson et al.
patent: 5475794 (1995-12-01), Mashiko
Ohmi Tadahiro
Shibata Tadashi
LandOfFree
Semiconductor neural circuit device does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Semiconductor neural circuit device, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Semiconductor neural circuit device will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2336774