Neural network processing system using semiconductor memories

BACKGROUND OF THE INVENTION
The present invention relates to a data processing system having a memory packaged therein for realizing a large-scale and fast parallel distributed processing and, more specifically, for realizing a neural network processing system.
Parallel distributed data processing using neural networks, called the "neuro-computing" (as will be shortly referred to as the "neural network processing") is noted in the field of acoustics, speech and image processing, and is described on pp. 145-168, "Parallel networks that learn to pronounce English test. Complex Systems 1 by Sejnowski, T. J., and Rosenberg, C. R. 1987, and "Neural Network Processing" published by Sangyo Tosho and edited by Hideki Asou. In neural network processing, a number of processing elements called "neurons" are connected in a network and exchange data through transfer lines called "connections" for high-grade data processing. In each neuron, the data (i.e., the outputs of the neurons) sent from another neuron are subjected to simple processes such as multiplications or summations. Since processing in the individual neurons and processing of different neurons can be carried out in parallel, the neural network processing is advantageous in principle because it offers fast data processing. Since algorithms for setting the connection weights of the neurons for a desired data processing have been proposed, data processing can be varied for the objects, as described on pp. 533-536, "Learning representations by back-propagation errors.", Nature 323-9 (1986) by Rumelhart, D. E. Hinton, G. E. and Williams, R. J., and on 2nd Section of "Neural Network Processing" published by Sangyo Tosho and edited by Hideki Asou.
SUMMARY OF THE INVENTION
First, the operating principle of the neural network will be described in connection with two representative kinds: the multi-layered network and the Hopfield network. FIG. 2(a) shows the structure of the multi-layered network, and FIG. 3(a) shows the structure of the Hopfield network. Both of these networks are constructed using connections of neurons. Here, the terminology "neurons" means "nodes" or "processing elements", whatever the case may be. The directions of the connecting arrows indicate the directions in which neuron outputs are transferred. In the multi-layered network, as shown in FIG. 2)a), the neurons are stacked in multiple layers so that the neuron outputs are transmitted in the direction from the input to output layers only. Input signals IN.sub.l, - - - , and IN.sub.n are input to the input layer, and output signals OUT.sub.l, - - - , and OUT.sub.n are output from the output layer. In the Hopfield network, on the other hand, the neuron outputs are fed back to an identical neuron and are transferred in two directions between two arbitrary neurons. The feedback to the identical neuron may be omitted.
FIGS. 2(b) and FIG. 3(b) show the processing principle to be accomplished in the neurons. This processing principle is similar in any network and will be described in connection with the multi-layered network with reference to FIG. 2(b). FIG. 2(b) shows a j-th neuron in the (S+1)th layer in an enlarged scale. This neuron is fed through the connection with the output values V.sub.ls, - - - , V.sub.is, - - - , and V.sub.nss of the neurons in the preceding layer, i.e., the S-th layer. Here, letters NS indicate the number of neurons in the S-th layer. In the neuron, the input products V.sub.ls T.sup.s.sub.ji, - - - , V.sub.is T.sup.s.sub.ji, - - - , V.sub.is, - - - , and V.sub.nss T.sup.s.sub.jns of the output values V.sub.ls, - - - , V.sub.is, - - - , and V.sub.ness of the neurons and the connection weights T.sup.s.sub.ji and so on, are calculated by means of a multiplier MT. Next, the sum of these products and an offset .theta..sub.js+l is calculated by means of an adder ADD. The offset .theta..sub.js+l may be omitted, as the case may be. Moreover, the result is input to a nonlinear transfer function circuit D in order to obtain the output value V.sub.js+ of the neurons. The nonlinear transfer function circuit D has characteristics, as shown in FIGS. 2(c) or 2(d), and outputs g(x) for an input x. FIG. 2(c) shows an example of the nonlinear transfer function which results in a binary output g1 or g2 depending upon whether or not the input x exceeds a predetermined threshold value xth. FIG. 2(d) shows an example using a sigmoid function for issuing continuous outputs. The nonlinear transfer function circuit D is given other characteristics, if necessary and may even be given linear characteristics.
The processing principle described above is also similar in the Hopfield network, as shown in FIG. 3(b). In the Hopfield network, however, not only the output of the neuron of the preceding layer but also the outputs of all neurons are input to one neuron. In the multi-layered network, as seen from FIGS. 2(a) and 2(b), one process is ended by first feeding the output values of the neurons of the input layer and then by updating the output values of the neurons in the next layer and then by updating the output values of the neurons of the next layer, i.e. the secondary layer. In the Hopfield network of FIG. 3(a), on the other hand, the output values of the individual neurons can be updated at suitable times because of the lack of any layer. In this Hopfield network, all the neuron output values are suitably given, and they are updated until they come to an equilibrium state. In order to distinguish the networks, the one in which the output values of all neurons are simultaneously updated, is called the "synchronized Hopfield network", whereas the one in which the output values are updated at arbitrary timings, is called the "unsynchronized Hopfield network".
One method used for realizing the above mentioned neural networks has employed software whereas the other has employed hardware. According to the method employing software, the processing of the neurons is carried out with a program written in computer language so that the number or structure of the neurons can be easily changed. Since, however, the processing is sequentially performed, the former method is disadvantageous in that the data processing time is abruptly elongated for an increased number of neurons. In the Hopfield network using an n number of neurons, an n times of products have to be calculated to update the output of one neuron. In order to update the output values of all neurons at least once, therefore, an n.sup.2 times of products must be calculated. In other words, the number of calculations will increase on the order of n.sup.2 with an increase in the neuron number n. As a result, the data processing time will increase on the order of n.sup.2 if multiplications are sequentially accomplished.
According to the method employing hardware, the processing time can be shortened by putting the neurons to be multiplied into the hardware. Another trial for speeding up processing has been made by executing the processing in parallel with a number of hardware neurons. If, however, the number of neurons is enlarged, the number of wiring lines acting as the signal lines between the neurons will increase the order of n.sup.2, thus making it difficult to realize a large-scale network.
The method of solving the wiring problem is exemplified on pp. 123-129 of Nikkei Microdevice, March, 1989, and will be described in principle in FIG. 4. FIG. 4 shows an example, in which a multi-layered network is composed of three layers, each having three neurons and is constructed of analog neuro-processors ANP and SRAM. The ANP is made by integrating one multiplier MT and one adder ADD of FIG. 2(b) and a nonlinear transfer function circuit D onto one chip. Another chip SRAM is stored with the connection weight belonging to each neuron. The neurons of different layers are connected through one signal line called the "analog common bus". Since the neuron output value of an input layer is input from the outside, the ANP and SRAM corresponding to the neurons of the input layer are omitted.
The operation are as follows. First, the connection weight necessary for the desired data processing for each ANP is read in each SRAM from the outside. Next, an input analog signal corresponding to the output value of one neuron in the input layer is input. The input signals are input in parallel directly to the ANPs corresponding to the neurons of the middle layer. The weight data are read out from the SRAM to each ANP in synchronism with the input signal. Next, the product of two signals is calculated, and the result is stored in each ANP. Next, the input analog signals corresponding to the output values of other neurons of the input layer are input, and their product is calculated. The calculation result is added to the value which is stored in each ANP of the middle layer. After similar calculations have been accomplished for the input analog signal corresponding to the output value of the last neuron of the input layer, the neuron output values V.sub.12, V.sub.22 and V.sub.32 of the middle layer are output consecutively to the analog common bus of the middle layer by the nonlinear transfer function circuit in the ANP so that similar calculations are continued. Finally, the neuron output values V.sub.13, V.sub.23 and V.sub.33 of the output layer are output consecutively to the analog common bus of the output layer by the nonlinear transfer function circuit in the ANP of the output layer.
According to the example of the prior art described above with reference to FIG. 4, the wiring problem can be avoided by driving the common bus in the time sharing manner. Since, moreover, the multiplications of the number of the neurons in one layer can be carried out in parallel, the data processing rate can be drastically raised, as compared with the method utilizing software since hardware processing is faster.
Since, however, the ANPs and the SRAMs are on different chips, it is difficult to realize a high density, large-scale network. Even if thirty two neurons could be integrated over a square chip of 10 mm, as described on pp. 129 of the above-specified Nikkei Micro Device, March, 1989, one adder, one multiplier and one nonlinear transfer function circuit need to be prepared for each neuron. Hence, it is difficult to integrate several hundreds or thousands of neurons on one chip.
Moreover, the examples of the prior art described above are accompanied by the following problems when they are applied to practices. An application of the multi-layered network is exemplified by the output of pronunciations and accents of English letters input to a three-layered network, as described on pp. 145-168 of Complex Systems 1 (1987) entitled "Parallel networks that learn to pronounce English text" and edited by Sejnowski, T. J. and Rosenberg, C. R. Seven English letters are encoded as the neuron output values of the first layer, an the codes corresponding to the pronunciations and accents of the central one of the aforementioned seven letters are output as the output values of the twenty six neurons of the third layer. In this example, depending upon the inputs, the output values of the neurons of the output layer may fail to coincide with the codes of the predefined pronunciations and accents but may given fuzzy values It is, therefore, necessary, to find the codes closet to those of all the pronunciations and accents and compare them to make them the correct answer. These comparisons between the output values and the desired values (i.e., the codes of the pronunciations and is accents in the aforementioned example) of the outputs are likewise required for pattern recognition by the neural network. This point is not taken into consideration in the aforementioned examples of the prior art so that the examples are not applicable when applied to the practical problem.
In the above mentioned examples of the prior art, moveover, the connection weight necessary for the desired data processing is determined by an external computer, and the result is written into the SRAMs of FIG. 4. As a result, the learning is wholly performed by software which makes it difficult to carry out at high speed.
In order to solve the problems thus far described, according to the present invention, there is provided neural network processing system which comprises: a memory for storing neuron output values, connection weights, the desired values of the outputs, and data necessary for learning; an input/output circuit for writing or reading data in or out of said memory; an arithmetic circuit for performing a processing for determining the neuron outputs such as the product, sum nonlinear conversion of the data stored in said memory, a comparison of the output value and its desired value, and a processing necessary for learning; and a control circuit for controlling the operations of said memory, said input/output circuit and said arithmetic circuit. The arithmetic circuit is constructed to include at least one adder, a multiplier, a nonlinear transfer function circuit and a comparator so that at least a portion of the processing necessary for determining the neuron output values such as the product or sum may be accomplished in parallel. Moreover, these circuits are shared among a plurality of neurons and are operated in a time sharing manner to determine the plurality of neuron output values. Still, mover, the above mentioned comparator compares the neuron output value determined and the desired value of the output in parallel.
Since the adder, the multiplier and the nonlinear transfer function circuit are shared among the plurality of neurons, as described above, the system for calculating the neuron output values in the neural network using numerous neurons can be realized by using a small number of circuits. Since, moreover, at least a portion of the neuron processing, such as the product and sum, is performed in parallel by the aforementioned arithmetic circuit, the data processing can be accomplished at high speed. Furthermore, since the output value obtained and the desired value of the output can be compared in parallel by the aforementioned comparator, the distance (i.e., the similarity between the obtained output value and the expected value, e.g., the hamming distance) between the obtained output value and the desired value can be determined at high speed. Since, furthermore, at least a portion of the processing necessary for the learning is accomplished by the hardware of the system, the learning speed can be increased.
An object of the present invention is to solve the problems owned by the aforementioned prior art.
Another object of the present invention is to provide a system for carrying out the calculations of the neuron output values at high speed with a small number of circuits in the network containing a number of neurons.
Still another object of the present invention is to give the aforementioned system a function to compare the neuron output value and the desired value at high speed.
A further object of the present invention is to give the aforementioned system a function to process at least a portion of the processing necessary for the learning.

These and other objects and many of the attendant advantages of the present invention will be readily appreciated as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 shows one embodiment of the structure, in which the data processing system according to the present invention is embodied over a single semiconductor chip;
FIG. 2(a) and 2(b) are diagrams showing the principle of the multi-layered neural network;
FIGS. 2(c) and 2(d) are diagrams showing the examples of the characteristics of the nonlinear transfer function circuit D;
FIGS. 3(a) and 3(b) are diagrams showing the principle of the Hopfield neural network;
FIG. 4 shows an example of the neural network processing system using the plural chips according to the prior art;
FIG. 5(a) and 5(b) show embodiments of the method for parallel processing of the multi-layered neural network and the Hopfield neural network, respectively;
FIGS. 6(a) and 6(b) show second embodiments of the method for parallel processing of the multi-layered neural network and the Hopfield neural network, respectively;
FIG. 7(a) shows one embodiment of the structure, in which the data processing system according to the present invention is realized by using a memory array capable of reading out a number of data on data lines by selecting one word line;
FIG. 7(b) shows one embodiment of the structure, in which the data processing system according to the present invention is realized by using two memory arrays capable of reading out a number of data on data lines be selecting one word line;
FIG. 8(a) shows one embodiment of the corresponding relations of memory cells to the neuron output values and the connection weights, in the case where the multi-layered neural network is realized by using the embodiment of FIG. 7(b), that is to say, the embodiment in which the neuron output value and the connection weight are realized with each memory cell;
FIG. 8(b) shows one embodiment of the characteristics of the nonlinear transfer function circuit D which is suitable in the case where binary memory cells are to be used in the embodiment of FIG. 8(a);
FIG. 8(c) shows on embodiment of the characteristics of the nonlinear transfer function circuit D which is suitable in the case where binary memory cells are used in the embodiment of FIG. 8(a);
FIG. 8(d) shows one embodiment of the method of selecting the word lines and the memory cells in the arithmetic mode in the embodiment of FIG. 8(a);
FIG. 9(a) shows a second embodiment of the corresponding relations of the memory cells to the neuron output values and the connection weights, in the case where the multi-layered neural network is realized by using the embodiment of FIG. 7(b), that is to say, the embodiment in which the neuron output values and the connection weights are realized using a plurality of memory cells;
FIG. 9(b) shows one embodiment of the characteristics of the nonlinear transfer function circuit D which is suitable in the case where the neuron output values and the connection weights are equivalently expressed by a plurality of memory cells in the embodiment of FIG. 9(a);
FIG. 9(c) shows one embodiment of a method of setting the values Xth1, - - - , and Xthp of FIG. 9(b) for giving the nonlinear function circuit D the sigmoid characteristics;
FIG. 10(a) shows a third embodiment of the corresponding relations of the memory cells to the neuron output values and the connection weights, in the case where the multi-layered neural network is realized by using the embodiment of FIG. 7(b), that is to say, the embodiment in which the neuron output values and the connection weights are realized using a plurality of memory cells;
FIG. 10(b) shows one embodiment of the characteristics of the nonlinear transfer function circuit D which is suitable in the case where the neuron output values and the connection weights are binarily expressed by a plurality of memory cells in the embodiment of FIG. 10(a);
FIG. 11 shows on embodiment in the case where a DRAM cell is used as the memory cell of the embodiment of FIG. 8(a);
FIG. 12 shows on embodiment of the relations between the arithmetic modes and the external signals;
FIG. 13(a) shows one embodiment of the operation waveforms in the memory mode of the embodiment of FIG. 11;
FIG. 13(b) shows on embodiment of the operation waveforms in the arithmetic mode of the embodiment of FIG. 11;
FIG. 14(a) shows one embodiment of the nonlinear transfer function circuit D which is suitable for the embodiment of FIG. 11 or the like;
FIG. 14(b) shows one embodiment of the characteristics of the nonlinear transfer function circuit D of FIG. 14(a);
FIG. 14(c) shows on embodiment of the input/output circuit 10 which is suitable for the embodiment of FIG. 11 of the like;
FIG. 14(d) shows one embodiment of the potential relation between the product and sum output line NO and the write line IA in the case where the nonlinear transfer function circuit of FIG. 14(a) and the write circuits of FIG. 14(c) are used;
FIG. 14(e) shows one embodiment of the read circuit OUT which is suitable for the embodiment of FIG. 11 of the like;
FIG. 14(f) shows one embodiment of the read latch circuit OUTLT which is suitable for the embodiment of the read circuit OUT of FIG. 14(e);
FIG. 15 shows a second embodiment of the multiplier MT in FIG. 11;
FIGS. 16(a) and 16(b) show examples of the SRAM cell;
FIG. 16(c) shows one embodiment in the case where the SRAM cells of FIGS. 16(a) and 16(b) are used as the memory cell of FIG. 8(a);
FIG. 17(a) shows one embodiment of the operation waveforms in the memory mode of the embodiment of FIG. 16(c);
FIG. 17(b) shows one embodiment of the operation waveforms in the arithmetic mode of the embodiment of FIG. 16(c);
FIG. 18(a) shows one embodiment in the case where the DRAM cells are used in the embodiment of FIG. 9(a) or the embodiment of FIG. 10(a);
FIG. 18(b) shows one embodiment of the structure of the block BLK1 in the embodiment of FIG. 18(a);
FIG. 18(c) shows one embodiment of the structure of the block BLK2 and the nonlinear transfer function circuit D in the embodiment of FIG. 18(a);
FIG. 18(d) shows one embodiment of the structure of the individual nonlinear transfer function circuits DSx (x=1, 2, - - - , and p) composing the nonlinear transfer function circuit D10 in the embodiment of FIG. 18(c);
FIG. 18(e) shows one embodiment of the characteristics of the individual nonlinear transfer function circuit DSx (x=1, 2, - - - , and p) composing the nonlinear transfer function circuit D10 in the embodiment of FIG. 18(c);
FIG. 19(a) shows on embodiment of the structure of the nonlinear transfer function circuit D10 which is suitable in case the neuron output values and the connection weights are binarily expressed in a plurality of memory cells in the embodiment of FIG. 18(a);
FIG. 19(b) shows one embodiment of the characteristics of the nonlinear transfer function circuit DSx (x=1, 2, - - - , and z) in the embodiment of FIG. 18(a);
FIG. 19(c) shows one embodiment of the characteristics in the embodiment of FIG. 19(a);
FIG. 19(d) shows one embodiment of the structure of the encoder in the embodiment of FIG. 19(a);
FIG. 20(a) shows one embodiment of the corresponding relations of memory cells to the neuron output values and the connection weights, in case the unsynchronized Hopfield neural network is realized by using the embodiment of FIG. 7(b), that is to say, the embodiment in which the neuron output value and the connection weight are realized in each memory cell;
FIG. 20(b) shows one embodiment of the corresponding relations of memory cells to the neuron output values and the connection weights, in case the synchronized Hopfield neural network is realized by using the embodiment of FIG. 7(b), that is to say, the embodiment in which the neuron output value and the connection weight are realized in each memory cell;
FIG. 21(a) shows one embodiment of the corresponding relations of memory cells to the neuron output values and the connection weights, in the case where the unsynchronized Hopfield neural network is realized by using the embodiment of FIG. 7(b), that is to say, the embodiment in which the neuron output value and the connection weight are realized with a plurality of memory cells;
FIG. 21(b) shows one embodiment of the corresponding relations of memory cells to the neuron output values and the connection weights, in the case where the synchronized Hopfield neural network is realized by using the embodiment of FIG. 7(b), that is to say, the embodiment in which the neuron output value and the connection weight are realized with a plurality of memory cells;
FIG. 22 shows one embodiment in the case where the neuron output values and the connection weights are enabled to take positive and negative values by using coding bits;
FIG. 23 shows one embodiment in the case where the system according to the present invention is given a function to compare the neuron output values and the desired values;
FIG. 24 shows one embodiment of the comparator for comparing the data read out to a plurality of data line pairs of the memory cells array TG and the memory cell array A to calculate the extend of similarity of the data; and
FIG. 25 shows one embodiment in which the neuron output values are updated faster by providing a register.

DETAILED DESCRIPTION OF THE INVENTION
FIG. 1 shows one embodiment in the case where the data processing system according to the present invention is integrated over a semiconductor chip.
The items integrated over a semiconductor chip (CHIP) include: a memory (A) for storing data; an input/output circuit (I/O) for performing at least one of the writing operation and reading the data in and from said memory; an arithmetic circuit (ARTM) for performing the arithmetic for determining neuron output values, the comparison (i.e., the similarity of the obtained output values and the desired values, e.g., the calculation of the hamming distance) of the output values and the desired values or the arithmetic necessary for learning by using the data stored in said memory; and a control circuit (CNT) for controlling the operations of the said memory, said input/output circuit and said arithmetic circuit. Each of the buses (BUS1, BUS2, BUS3, BUS4, etc.) connecting the individual blocks is made of not only one wiring line but also a necessary number of wiring lines. The aforementioned memory can be stored with the connection weights and the neuron output values necessary for neural network processing the desired values of the outputs or the data necessary for the learning. According to the present embodiment, nonlinear network processing such as the calculations of the neuron output values, the comparisons of the output values with the desired values, or the calculations necessary for learning can be performed in the following ways.
First, the method of calculating the neuron output values will be described. In the first step, the connection weights necessary for the calculations for neural network processing, and the neuron output values, or the offsets are read out in parallel from the memory to the arithmetic circuit through a bus 1. Next, the calculations such as the product and sum or the nonlinear transfer necessary for determining the neuron output values are accomplished by the arithmetic circuit, and the obtained results are written in the memory through the input/output circuit. The operations described above are continued the necessary number of times to determine the neuron output values. The arithmetic circuit may either determine one of a plurality neuron output values by a single operation or perform a portion of calculations for determining the neuron output values Thus, the data processing can be accomplished by the various networks such as the multi-layered network or the synchronized or unsynchronized Hopfield network. Incidentally, in order to update the output values of all neurons synchronously, the synchronized Hopfield network needs to be stored with the output values of all the neurons until all the output values of all the neurons have been updated. In this case, the output values of all the previous neurons may be stored in the memory so that they may be used for updating the output values of the neurons.
According to the present embodiment, a desired number of multipliers, address and nonlinear transfer function circuits necessary for calculating the neuron output values may be provided in the arithmetic circuit so that they may be used repeatedly. This makes it possible to make the number of circuits far smaller that the case in which those circuits are prepared for each of the neurons. The example of the prior art of FIG. 4 requires two hundreds multipliers, adders and nonlinear transfer function circuits for realizing the multi-layered network having three layers each composed of one hundred neurons. In the present embodiment, on the contrary, it is sufficient to prepare at least one multiplier, at least one adder and at least one nonlinear transfer function circuit. Even if the multiplications necessary for updating one neuron output value would be accomplished in parallel, it would be sufficient to prepare one hundred multipliers, one adder and one nonlinear transfer function circuit. According to the present embodiment, therefore, the number of circuits can be drastically reduced, as compared with that of the prior art. Incidentally, the above-specified difference will become larger for larger scale networks. Similar situations will apply to another network such as the Hopfield network.
Not on the calculation speed such as multiplication, but also the amount of calculations to be carried out in parallel makes a significant contribution to data processing speed where the neuron output values are to be determined. In the Hopfield network using n number of neurons, for example, n.sup.2 products must be calculated for updating the output values of all the neurons, as has been described earlier. If the multiplications are accomplished sequentially, therefore, the updating of the output values of all the neurons takes at least a time period of n.sup.2 times as long as that required for one multiplication. As a result, the time period required for the multiplications will abruptly increase on the order of the square of the neuron number with an increase in the number of neurons. Similar circumstances will also apply to the multi-layered network. This makes it desirable to calculate multiple multiplications in parallel. Next, an example of the arithmetic system for raising the data processing speed by making the multiplications in parallel so as to determine the neuron output values in the embodiment of FIG. 1 will be described.
FIG. 5 illustrates the multi-layered network at (a) and the Hopfield network at (b) on one system for parallel computations. In the present embodiment, the products necessary for determining one neuron output value are calculated, as shown. Specifically, the output value of the neuron of the preceding layer, which is input to one neuron, and the connection weight for the said output value of the neuron under consideration are read out in parallel from the memory, and their products are calculated in parallel. Thus, the time period required for the multiplications will increase on the order of the neuron number with an increase in the neuron number. As a result, the data processing time can be drastically shortened, as compared with the case in which the multiplications are accomplished sequentially. In FIG. 5, only the multiplications necessary for updating the output value of one neuron are executed in parallel. However, the embodiment of FIG. 1 should not be limited thereto but may naturally add the arithmetic circuits within a range allowed by the degree of integration, to update the output values of the plurality of neurons in parallel. In this case, data processing can be accomplished at higher speed. In addition, the parallel calculations can also be accomplished by another system, as shown in FIGS. 6(a) and 6(b).
FIG. 6 shows one embodiment, in which the multiplications are executed in parallel for a plurality of neurons to be fed with the output value of one neuron in the multi-layered network of FIG. 6(a) and in the Hopfield network of FIG. 6(b). In this method, the neural output values and the connection weights are read out from the memory, and the calculations necessary for updating the neuron output values are executed bit by bit for the plurality of neurons. This makes it impossible to realize the unsynchronized Hopfield network. Since, however, the time period required for the multiplications will increase on the order of the neuron number with the increase in the neuron number like the system of FIG. 5, the data processing time can be drastically shortened, as compared with the case in which the multiplications are carried out sequentially.
In the example of the prior art of FIG. 4, too, the calculations are executed in parallel. As will be described below, however, the structure of FIG. 1 can be realized with a smaller number of circuits than that of the example in the prior art. In the systems shown in FIGS. 6(a) and 6(b), only one multiplier operates in parallel in each neuron, as hatched. In the embodiment of FIG. 1, therefore, the arithmetic circuit may be provided with multipliers in a number equal to that of the neurons to be calculated at one time. This is done so that the system can be realized with a smaller number of circuits than that in the case of the prior art in which multipliers are provided for all the individual neurons. In the multi-layered network having three layers, each composed of three neurons, for example, a similar parallelism can be realized by suing three individual multipliers, adders and nonlinear transfer function circuits according to the embodiment of FIG. 1. This is contrary to the case of the prior art which is equipped with six individual multipliers, adders and nonlinear transfer function circuits.
Thus, according to the embodiment shown in FIG. 1, a system for data processing similar to that of neural networks using numerous neurons can be realized with the minimum number of necessary circuits by sharing the adders, multipliers and nonlinear transfer function circuits of the arithmetic circuit among the plurality of neurons. By executing the calculations such as the products or sums with the aforementioned arithmetic circuit, moveover, the data processing can be accomplished at high speed. Incidentally, with parallel calculations, the number of wiring lines between the memory and the arithmetic circuit has to be increased to send many data at once to the arithmetic circuit. In FIG. 1, however, the memories and the arithmetic devices are arranged over a common chip so the number of wiring lines on the bus can be easily increased.
Although the method of calculating the neuron output values has been described hereinbefore, a neuron output value and its desired value can be compared according to the embodiment of FIG. 1. For this comparison, the desired value may be stored in advance in the memory so that its distance from the output value obtained by the aforementioned method may be calculated by the arithmetic circuit. This operation is to calculate the similarity between the desired value and the calculated value. At this time, the desired,value and the output value, composed of numerous bits, can be simultaneously read out to the arithmetic circuit and processed in parallel with ease by increasing the number of wiring lines on the bus 1. Thus, according to the embodiment of FIG. 1, data processing such as pattern recognition can be executed at high speed, as compared with the case in which the comparison is accomplished serially bit by but by using an external computer.
According to the embodiment of FIG. 1, moreover, the learning can be accomplished at higher speed than that of the case using software, by executing the calculations necessary for learning with the arithmetic circuit. This specific embodiment will be described below.
Neural network processing is advantageous in that it can process various data by changing the connection weights. This advantage can be easily exploited according to the embodiment of FIG. 1 by rewriting the connection weight stored in the memory. Moreover, several kinds of connection weights necessary for different data processings can be stored in advance by making the capacity of the memory larger than that necessary for calculating the neuron output values. In this case, the benefit that different kinds of data can be continuously processed without losing the time period for rewriting the connection weights can be realized. In addition, in the case where numerical input data are to be continuously processed, the necessary input data or the obtained data can be stored in advance in a portion of the memory. Thus, the frequency for switching the reading, calculating and output modes can be reduced to shorten the processing time, as compared with the case in which the operations of reading each input data in the memory and calculating and outputting it externally are repeated.
Next, a more specific embodiment, based upon the embodiment of FIG. 1, will be described. For simplicity, the case where the arithmetic circuit is given the function of calculating the neuron output values will be described at first, and the method of giving the comparing or learning function will be described later.
FIG. 7(a) shows one embodiment in the case where a lattice-shaped memory cell array is used in the memory of the embodiment of FIG. 1. In FIG. 7(a), letter A designates a memory cell array which is composed of: a plurality of data lines (D); a plurality of word lines (W) arrayed to intersect the data lines (D); and the memory cells (MC) arrayed at the desired intersections. As a result, the signals of the different memory cells can be read out onto the plurality of data lines by selecting one of the word lines. Numeral 12 designates an arithmetic circuit (ARTM). Numerals 10, 11, 13, 14, 15 and 16 designate circuits corresponding to the control circuit (CNT) of FIG. 1. The numerals 10 and 15 designate address buffers for X-addresses and Y-addresses, and the numerals 11 and 14 designate a decoder and a driver for X-addresses and Y-addresses, respectively. Numeral 13 designates an array control circuit for controlling the memory cell array. Numeral 16 designates a clock generator for generating clocks to control the operations of the memories on the basis of the signals input from the outside. Letters OUT and WR designate a read circuit and write circuit, respectively. A chip select CS is a chip selection signal. A write control signal WE is a signal for switching the write and read operations for establishing the write operation at a low level and the read operation at a high level. Letters NE designate an arithmetic circuit control signal for starting the arithmetic circuit at a low level and interrupting the same at a high level to act as an ordinary memory. In the following, the state of the signal NE at the high level will be called with "memory mode", and the state at the low level will be called the "arithmetic model". In the memory mode, a desired memory cell is selected according to the X-address and the Y-address so that a write data DI can be written in that cell or so that data can be read out from the same cell and output as a read data DO. In the arithmetic mode, the data stored in the memory cell is read out to the arithmetic circuit 12 so that the arithmetic result of the arithmetic circuit 12 or the data according to the arithmetic result can be written in the memory cell through the input circuit. By selecting one word line, according to the present embodiment, the data of all the memory cells on the selected word are output to the data lines. As a result, numerous data can be easily latched in the arithmetic circuit 12 so that many calculations can be accomplished in parallel. In order to calculate the neuron output values according to the present embodiment, the mode is first set to the memory mode to stop the arithmetic circuit, and the necessary connection weight, neuron output value (i.e., the input signal at first), offset and so on are written in the memory. Next, the mode is set to the arithmetic mode to start the arithmetic circuit reading the necessary data out to the arithmetic circuit by selecting one word line. Next, the result is written in the memory circuit. If the read of the data necessary for the calculations and the write of the result are further continued the necessary number of times, neural network processing can be accomplished at high speed. As has been described above, according to the embodiment shown in FIG. 7(a), many data can be written at once in the arithmetic circuit since the embodiment is suited for parallel calculations of the type shown in FIG. 5 or FIG. 6. Thus, according to the present embodiment, parallel calculations make it possible to execute neural network processing at high speed. By using the arithmetic circuit 12 repeatedly, moreover, the plurality of neurons can share the output value with the calculating circuit to easily effect a high degree of integration.
In the case, on the other hand, where parallel calculations are to be accomplished by using the data stored in the memory cells on the plurality of word lines, a primary storage register can be provided in the arithmetic circuit so that it may store one time, the data obtained by selecting the word lines and execute the calculations of the stored data together with the data read out by selecting other word lines.
As in the embodiment shown in FIG. 7(b), moreover, two memories, A and B can be provided. In FIG. 7(b), characters 13A and 13B designate array control circuits for controlling the memory cell arrays A and B, respectively. Other circuits such as a decoder is not shown in FIG. 7(b). According to the structure of FIG. 7(b), the memory cell data on the two word lines of memory cell arrays A and B can be written in the arithmetic circuit by selecting one word line each of the memory cell arrays A and B. by using the structure of FIG. 7(b) the controls of the reading or writing operations can be simplified because the memory arrays can be used separately according to the kinds of data such that memory cell array A can store the neuron output value whereas memory cell array B can store the connection weight. Incidentally, in the embodiments of FIGS. 7(a) and 7(b), the write data DI and the read data DO may be processed in plurality and in parallel, or arrays A and B may be provided separately with the read circuit OUT and the write circuit WR.
In the embodiments of FIGS. 7(a) and 7(b), the selection of a specific memory cell can be accomplished like the ordinary memory according to the address. By changing the order of selecting the address, therefore, those embodiments can be flexibly applied to the various networks or various parallel arithmetic systems.
In the embodiments of FIGS. 7(a) and 7(b), the memory can be exemplified by a highly integrated semiconductor memory such as the SRM or SRAM. In this case, the memory can store many data so that a large-scale network can be integrated into one chip.
Next, the method of realizing the multi-layered network by using the structure of FIG. 7(b) will be described below in detail. The parallel arithmetic system is exemplified by taking the system of FIG. 5(a). It is assumed that the number of layers be m and that the number of neurons in each layer be n. Incidentally, the offset .theta. of each neuron, as shown in FIG. 2(b) or FIG. 3(b), will be omitted here to simplify the description. As is apparent from FIG. 2(b) or FIG. 3(b), however, the offset .theta. of each neuron can be handled like the output from another ordinary neuron. This may be accomplished by two methods. The first provides one neuron having an output value of 1 at all times to set the connection weight of it and each neuron to the offset .theta.. The second method increases the neuron output value, which is to be input to each neuron, by 1 in order to set the value of the offset .theta. of each neuron and the corresponding connection weight to 1 so that their product may be added to the total sum of the products of other neuron output values and the connection weights. FIG. 8(a) shows one embodiment in which the memory cells correspond to the connection weights and the neuron output values. Letter D designates the nonlinear transfer function circuit; characters c1, c2, - - - , and cn designate the adders; and characters m1, m2, - - - , and mn designate the multipliers, The adders c1, c2, - - - , cn together constitute the multi-input adder ADD of FIG. 2(b). The memory cell array A is stored with the neuron output values, and the memory cell array B is stored with the connection weights. Although what is shown in FIG. 8(a) is the memory cells for storing the neuron output values and the connection weights, it is quite natural that the memory cells may be stored with other data such as the offsets .theta. of the individual neurons or the data necessary for the learning, if necessary. As shown, the memory cells located at the intersections of the word lines s and the data lines i in the memory cell array A are stored with neuron output values V.sub.is. In other words, the output values of the neurons of the common layer are arranged on the common word line. In the memory cell array B, the memory cells located at the intersections between the word lines (s, j) and the data lines i are stored with connection weights T.sup.s.sub.ij.
FIGS. 8(b) and 8(c) show one embodiment of the input/output characteristics of the nonlinear transfer function circuit D. FIG. 8(b) shows the embodiment having binary outputs g1 and g2. Characters x1 and x2 indicate the lower limit and the upper limit of the input x, respectively. In FIG. 8(b), the output is g2, if the input x exceeds the threshold value xth, but otherwise it is g1. Therefore, the embodiment of FIG. 8(b) is suitable when the memory cell used is binary. FIG. 8(c) shows an embodiment having tetral outputs ga and gb between the outputs g1 and g2. The present embodiment is a suitable example when the tetral memory cells are used The gap between the elements g1, ga, gb and g2 can naturally be changed. If necessary, although they are shown equal distant in FIG. 8(c) In the case where the memory cells are exemplified by those capable of storing data having continuous values, i.e., the so-called "analog values", the nonlinear transfer function circuit D to be used may have the characteristics shown in FIG. 2(d).
FIG. 8(d) shows one embodiment of the correspondences in the embodiment of FIG. 8(a) between the word line selecting method for determining the neuron output values of the final layer from is the neuron output values of the input layer and the write destination addresses. The operations of FIG. 8(a) will be described in the following with reference to FIG. 8 (d). The neuron output values v.sub.11, V.sub.21, - - - , and V.sub.n1 of the input layer are written in advance in the input/output circuit (although omitted from FIG. 8) in the memory cells on the word line S=1 of array A. First, the word lines of (s, j)=(1, 1) of array B are selected simultaneously, although not necessarily completely. Then, the neuron output values V.sub.11, V.sub.21, - - - , and V.sub.ni of the input layer are output to the data lines of i=1, 2, - - - , and n of array A. On the other hand, the connection weights T.sup.111, T.sup.1.sub.12, - - - , and T.sup.1.sub.1n are output to the data lines of i=1, 2, - - - , and n of array B. These values are input to the multipliers m1, m2, - - - , and mn so that their products T.sup.1.sub.11 V.sub.11, T.sup.1.sub.122 V.sub.2l, - - - , and T.sup.1.sub.1n V.sub.n1 are input to the adders c1, c2, - - - , and cn. The results (T.sup.1.sub.11 V.sub.11 =T.sup.1.sub.12 V.sub.21, - - - , +T.sup.1.sub.1n V.sub.n1) are input to the nonlinear transfer function circuit D. The output of this nonlinear transfer function circuit D is written through the write circuit WR (although omitted) in the memory cells corresponding to the write destination addresses of (s, 1)=(2, 1) in array A. Thus, the value of the first neuron output value V.sub.21 of the second layer is calculated. Next, the word line of s=1 of array A and the word line of (s, j)=(1, 2) of array B are simultaneously selected. Then, the neuron output values V.sub.11, V.sub.21, - - - , and V.sub.n1 of the input layer are output to the data lines of i=1, 2, - - - , and n of array A. On the other hand, the connection weights T.sup.1.sub.21, T.sup.1.sub.22, - - - , and T.sup.1.sub.2n are output to the data lines of i=1, 2, - - - , and n of array B. These values are input to the multipliers m1, m2, - - - , and mn so that their products T.sup.1.sub.21 V.sub.11, T.sup.1.sub.21 V.sub.21, - - - , and T.sup.1.sub.2n V.sub.n1 are input to the adders C1, C2, - - - , and cn. The results (T.sup.1.sub.21 V.sub.11 +T.sup.1.sub.22 V.sub.21, - - - , +T.sup.1.sub.2n V.sub.n1) are input to the nonlinear transfer function circuit D. The output of this nonlinear transfer function circuit is written through the write circuit (although no shown) in the memory cells corresponding to the write destination addresses of (s, i)=(2, 2) in array A. Thus, the value of the second neuron output value V.sub.22 of the second layer is calculated. All the neuron output values can be calculated by continuing the operations thus far described according to FIG. 8(d). According to the present embodiment, one neuron output value can be determined by executing the read and write operations once in the arithmetic mode so that the neural network processing can be accomplished at high speed. Moreover, since the arithmetic circuit can be shared among all the neurons, high integration can be attained. Incidentally, FIG. 8(d) shows only one example of the assignment of the memory cells, and the present invention should not be limited thereto but may be modified in various manners. For example, the plurality of input data may be continuously processed, as has been described above. In this case, a plurality of sets of neuron output values of an input layer are required. For this operation, the neuron output values of the input layer corresponding to the plurality of input data may be written in advance on the plurality of different word lines of array A so that they may be used consecutively. Thus, the neuron output values of the input layer need not be read in for each data processing so that the data processings can be continuously accomplished at high speeds.
In this example, one memory cell is used for storing the neuron output value and the connection weight. This allows only binary values to be taken as the neuron output values and the connection weights in the case where the binary memory cells are used. By using the multi-valued memory cells, as was described above, the neuron output values and the connection weight values could be increased, but the multi-valued memory cells may have their reliability decreased due to a problem with the S/N ratio. In this case, a plurality of memory cells may be used for storing the neuron output values and the connection weights, as will be described below.
FIG. 9(a) shows one embodiment of the case in which a number p of memory cells are used for storing one neuron output value and in which a number q of memory cells are used for storing one connection weight. The suffix i, j or s appearing in FIG. 9(a) to indicate the neuron output value or the connection weight corresponds to that of the embodiment shown in FIG. 8. In the embodiment of FIG. 9(a), the number p of continuous memory cells on one word line in array A expresses one neuron output value, and the number 1 of continuous memory cells on one word line in the array B expresses one connection weight.
The calculations of the neuron output values are carried out in the following manner. First, like the embodiment of FIG. 8, the word line of s=1 of the array A and the word line of (s, j)=(1, 1) of the array B are selected simultaneously. Then, to the data line group of i=1, 2, - - - , and n composed of p number data lines of array A, there is output the data expressing the neuron output values of V.sub.11, V.sub.21, - - - , and V.sub.n1 of the input layer, which are input group by group to the adders a1, a2, - - - , and an. On the other hand, the data expressing the connection weights T.sup.1.sub.11, T.sup.11.sub.12, - - - , T.sup.1.sub.1n which are read in group by group to the adders b1, b2 - - - , and b3 are output to the data line group of i=1, 2, - - - , and n which is composed of q number of data lines. The neuron output values V.sub.11, V.sub.21, - - - , and V.sub.n1 and the connection weights T.sup.1.sub.11, T.sup.1.sub.12, - - - , and T.sup.1.sub.1n are composed by the adders a1, a2, - - - , an and b1, b2, - - - , bn mentioned above and input, as shown, the multipliers m1, m2, - - - , and mn to produce the products T.sup.1.sub.11V.sub.11, T.sup.1 V.sub.21, - - - , and T.sup.1.sub.1n V.sub.n1. These products are input to the adders c1, c2, - - - , and cn so that their results (T.sup.1.sub.22 V.sub.11 +T.sup.1.sub.12 V.sub.21, - - - +T.sup.1.sub.1n V.sub.n1) are input to the nonlinear transfer function circuit D. The output of the nonlinear transfer function circuit is written through the write circuit WR (although not shown) to the p number of memory cell groups corresponding to the write destination address (s, i)=(2, 1) in array A The output values of all the neurons can be determined by continuing similar operations by using the same address as that of FIG. 8(d).
Since, according to the aforementioned embodiment, one neuron output value is expressed with p number of continuous memory cells on one word line in array A, the multi-valued neuron output values can be expressed by using the binary memory cells Since, moreover, one connection weight is expressed with q number of continuous memory cells on one word line in array B, the multi-valued connection weights can be expressed by using the binary memory cells. As a result, the multiple values, such as the neuron output values or connection weights, can be expressed using the binary memory cells. In the aforementioned embodiment, moreover, the frequency of switching the addresses is identical to that of the embodiment of FIG. 8 so that the data can be processed at higher speed like in the embodiment of FIG. 8. In order to write the result of the nonlinear transfer function circuit in p number of memory cells expressing the neuron output values, p number of writing operations may be continuously executed but can be easily accomplished in parallel by providing the p number of write circuits. In this case, it is possible to avoid the time losses for writing to the plurality of memory cells. By providing a plurality of read circuits, on the other hand, it is naturally possible to increase the speed of the reading operations. Incidentally, if a multiple input circuit is used for the multipliers m1, - - - , and mn in the embodiment of FIG. 9(a), a similar circuit can be realized without provided the adders a1, - - - , and an, and b1, - - - , and bn. In addition, the structure of the arithmetic circuit can be modified in various manners.
The embodiment of FIG. 9(a) uses p number of memory cells for storing the neuron output values and q number of memory cells for storing the connection weights. In other words, the neuron output values are expressed in p bits, and the connection weights are expressed in q bits. Since there are various methods of expressing the data in a plurality of bits, an expression method may be selected, if necessary, and it is accordingly sufficient to design the characteristics of the adders, the multipliers and the nonlinear transfer function circuits. For example, the neuron output values can be expressed with the number of memory cells of p bits expressing the neuron output values which have a content of 1. FIG. 9(b) shows one embodiment of the input/output characteristics of the nonlinear transfer function circuit D suited for such a case. In FIG. 9(b), characters g1, g2, - - - , and gp designate the p number of outputs of the nonlinear transfer function circuit D. Their individual outputs take the value 0 or 1, which is written in the p number of corresponding memory cells through the write circuit WR (although not shown). The outputs g1, g2, - - - , and gp take the value 1, when the individual inputs exceed the values xth1, xth2, - - - , and xthp, and otherwise take the value 0. These values xth1, xth2, - - - , and xthp may be at an equal or arbitrary distance between the upper limit x1 and the lower limit x2 of the inputs. If the distance between the values xthk and xthk+1 (k=1, - - - , and p-1) is set, as shown in FIG. 9(c), the nonlinear transfer function circuit g can be shown to have sigmoid characteristics. According to the present embodiment, one neuron output value can be given p number of values by the p number of memory cells. In the present embodiment, the p number of memory cells expressing the neuron output values are equivalently handled. Specifically, it does not matter for the influences upon the neuron output values if the p number of memory cells have its data inverted or fixed. As a result, the influences upon the neuron output values due to the fall of the data of one memory cell can be made lower than those of the general binary expressions. This expression method will be referred to as the "equivalent expression" in the following. So far the neuron output values have been described, but the aforementioned equivalent expression can naturally be used in connection weights.
The binary expression can naturally be used. This case is suited for expressing many values with less memory cells because the 2.sup.p values can be expressed in p bits. FIG. 10(a) shows one embodiment of the case, in which the binary expressions are used for the neuron output values and connection weights. Only the memory cells on the data line of i=h (h=1, 2, - - - , and n ) are shown in array A, and only the memory cells on the data line of i=h and on the word line of s=f (f=1, 2, - - - , and m-1) are shown in array B. In FIG. 10(a), letters WT designate a weighing circuit for weighing the signals of the memory cells and transmitting them to the address ah and bh. Here, the weighing coefficients are varied for every memory cell, as shown, from 1 to 2.sup.p for the neuron output values and from 1 to 2.sup.q for the connection weights. As a result, the neuron output values and the connection weights to be input to the multiplier mh can take 2.sup.p and 2.sup.q number of values, respectively. The address selecting method for calculating the neuron output values may follow FIG. 8(d) like the embodiment of FIG. 9. FIG. 10(b) shows one embodiment of the characteristics of the nonlinear transfer function circuit D in the embodiment of FIG. 10(a). The output g1 alternately repeats the values 0 and 1, each time the input varies by (x2-x1)/2.sup.p, and the output g2 alternately repeats the values 0 and 1 for a period twice as long as that of the output g1. The periods are likewise varied by times so that the output gp is set from 0 to 1 across the value of (x2-21)/2. In other words, the nonlinear transfer function circuit D may be so designed that it may operate as an A/D converter. In the present embodiment, too, the nonlinear transfer function circuit D can be so designed that the neuron output values may increase nonlinearly for the input. In order to increase the neuron output values according to the sigmoid function in response to the input, for example, the periods for the individual outputs may be decreased with the increase in the input, with the period ratios between the different values of g being held constant, so that the periods may be increased with the increase in the input when the input exceeds the value of (x2-21)/2. As has been described hereinbefore, according to the embodiments shown in FIGS. 10(a) and 10(b), the neuron output values and the connection weights can be expressed to have the 2.sup.p values and the 2.sup.q values, respectively, by using the p and q numbers of memory cells, respectively. Thus, the embodiments are suited for giving the neuron output values and the connection weights multiple values with less memory cells. Incidentally, in the present embodiment, too, various modifications can naturally be made such that the functions of the weight circuit WT and the adders a1, - - - , and an, and b1, - - - , and bn are given to the multipliers by using multiple input circuits with the multipliers m1, - - - , and mn. So far, the embodiments using the equivalent expressions and the binary notations have been described. In addition, there are various methods such as the method of expressing data in a plurality of bits, which can be separately used, if necessary.
Next, will be described an embodiment, in which a dynamic memory cell (i.e., DRAM cell) composed of one MOS transistor and one capacitor is used in the memory.
FIG. 11 shows an embodiment in which the embodiment shown in FIG. 8(a) is composed of the DRAM cells. In FIG. 11 array A and array B are comprised of: a plurality of intersecting data line pairs DA1 and DA1, - - - , and DAn and DAn, and DB1 and DB1, - - - , and DBn and DBn, and word lines WA1, WA2, - - - , and WAm, and WB1,1, WB1,2, - - - WB1,n, and WB2,1, WB2,2, - - - , and WBm-1,n; and memory cells MC disposed at their intersections. Each memory cell MC is arranged at the intersections between the paired data lines and the word line. Thus, the embodiment has the so-called "folded-bit line structure". In FIG. 11, letters PR, SA, RSA and WS designate a precharge circuit, a sense amplifier, a read sense amplifier, and a write switch respectively, and correspond to the array control circuits 13A and 13B of FIG. 7(b) Letters MT appearing in the arithmetic circuit designate a multiplier. Numeral 16 designates a clock generator for generating clocks .phi..sub.A and .phi..sub.B for controlling other circuits in response to addresses ADDA and ADDB and chip select signals CSA and CSB attained externally.
In the following, the operations of the embodiment of FIG. 11 will be described with reference to FIG. 12 and FIGS. 13(a) and 13(b). FIG. 12 shows one embodiment between the operation modes and the external signals. As has been described hereinbefore, the word lines of memory cells arrays A and B are selected one by one in the earlier half of the arithmetic mode, and one word line of memory cell array A is selected in the later half. In the memory mode, on the other hand, memory cell arrays A and B are independently subjected to the read and write operations. In FIG. 12, the operation modes are further divided so that those controls may be facilitated. Letters AR, AW, BR, BW in the memory mode distinguish the read mode on array A from the write mode in the array, and the read mode in array B from the write mode in the array B, respectively. On the other hand, letters NR and NW in the arithmetic mode designate the earlier half for reading and calculating the individual data and the later half for writing the arithmetic results, respectively. In order to switch those size modes, the present embodiment uses four external input signals: chip select signals CSA and CSB; write control signal WE; and arithmetic circuit control signal NE. The chip select signals CSA and CSB assign the selections of the chips and the arrays A and B. The chips are unselected if both the signals CSA and CSB are at the H (i.e., high) level; the array is selected if the signal CSA is at the L (i.e., low) level; and the array B is selected if the signal CSB is at the L (i.e., low) level. The write control signal WE is used for switching to write and read, as has been described hereinbefore, and establishes the reading operation at the H level and the writing operation at the L level. The signal NE also establishes the memory mode at the H level and the arithmetic mode at the L level, as has been described hereinbefore. If for example, both the signals CSA and CSB are at the L level and the signal WE is at the H level whereas the signal NE is at the L level, the earlier half mode NR of the arithmetic is established, in which both arrays A and B are read out. Since switching of arrays A and B is assigned by the chip select signal, the address signal can divide the addresses into the address group ADDA for selecting the memory cells of the array A and the address group ADDB for selecting the memory cells of the array B. Here, the address group ADDA is the generic name of the X-addresses for selecting the word lines of the array A and the Y-addresses for selecting the data lines of the array A. Likewise, the address group ADDB is the generic name of the X-addresses for selecting the word lines of the array B and the Y-addresses for selecting the data lines of the array B. In each operation mode, these address groups are applied to the address pins in accordance with FIG. 12. According to the embodiment of FIG. 12 thus far described, the two chip select signals are provided to switch arrays A and B thereby separating the addresses between arrays A and B. Since arrays A and B can be independently selected, it is possible to control each operation mode easily by selecting array A and/or B. Incidentally, the relation between the operation mode and the external signal can naturally be modified in various manners in addition to those of FIG. 12. For example, one can adopt: the method of adding addresses for switching the arrays A and B by using the chip select signal CS only; or the method of generating the X-address for selecting the word line of array B in the mode NR from the counter disposed in the chip, by not dividing the addresses for arrays A and B but selecting either array a or B only.
FIG. 13(a) shows an embodiment of the waveforms of FIG. 11 in the memory mode, and FIG. 13(b) shows an embodiment of the waveforms of FIG. 11 in the arithmetic mode.
The operation of the memory mode are similar to the read and write operations of the ordinary DRAM. FIG. 13(a) shows the voltage waveforms in the case where the read operations (in the mode AR) and the write operations (in the mode AW) are continuously executed for the memory cell at the intersection between the word line WA1 and the data line DA1 in array A in the memory mode. In FIG. 13(a), letters Vcc designate a positive supply potential. Since, in the memory mode, the arithmetic circuit control signal NE is at the high level, the arithmetic circuit start signal .phi. .sub.N is fixed at the low level so that the arithmetic circuit 12 is OFF. Before the start of the reading operation, signals PPA and PNA are set to Vcc/2 so that the sense amplifier SA is OFF. Since a precharge signal .phi..sub.PA is at the high potential, on the other hand, the precharge circuit PR is turned on to short the data line pairs DA1 and DA1, - - - , and DAn and DAn and to set the potential at a precharge potential VH. This precharge potential VH is set to Vcc/2 according to the so-called "half-precharge method". When the chip select signal CSA drops to the low potential, the precharge signal .phi..sub.PA falls to turn off the precharge circuit PR so that the word line WA1 selected by the address signal ADDA and a read Y-Select signal YRA1 are transited to the high potential. As a result, the MOS transistors of all the memory cells MC connected with the word line WAI are rendered conductive to individually establish delicate potential differences on the data line pairs DA1 and DA1, - - - , and DAn and DAn in accordance with the electric charges stored in the capacitor. This potential difference is read out and is detected by the sense amplifier RSA fed with the Y-Select signal YRA1 so that it is converted into the impedance difference of read lines OA and OA. This impedance difference is converted by the input/output circuit into a voltage difference, which is amplified so that the content of the memory cell, i.e., the potential corresponding to 1 or 0 is output as the read data DO. The so-called "rewriting operation" is executed in parallel with the aforementioned operations in the following manner. After the individual delicate potential differences have been established in the data line pairs DA1 and DAI, - - - , and DAn and DAn, the signal PPA is transited to the high potential whereas the signal PNA is transited to the low potential to start the sense amplifier SA. As a result, the delicate potential difference established in the data line pairs is amplified to transit the data lines at the high potential to the level Vcc and the data lines at the low potential to 0 V. As a result, the capacitors of all the memory cells MC connected with the word line WA1 are written again with the potential corresponding to the data before read. When the chip select signal CSA takes the high potential after the end of the rewriting operation, the selected word line WA1 and the read Y-select signal YRA1 are transited to the low potential, and the signals PPA and PNA are then transited to Vcc/2 to turn off the sense amplifier SA and transit the precharge signal .phi..sub.PA to the high potential. As a result, the data line pairs are shorted, and the precharge potential VH is set with the potential, until the initial state is restored. The operations thus far described are the reading operations.
Subsequently, the operations are shifted to the writing operations (in the mode AW) of the same cells. When, in the writing operations, the chip select signal CSA takes the low potential and the write control input WE takes the low potential, the data given to the write data DI are written in the selected memory cell in the array A. In the writing operation, too, the precharge signal .phi..sub.PA is dropped at first to turn off the precharge circuit PR when the chip select signal CSA drops to the low potential. Next, the word line WA1 selected by the address signal ADDA and the Y-Select signal YRA1 are transited to the high potential. As a result, the MOS transistors of all the memory cells MC connected with the word line WA1 are rendered conductive so that the delicate potential differences are established individually in the data line pairs DA1 and DA1,- - - , and DAn and DAn in accordance with the electric charges stored in the capacitor. The delicate potential differences established in the data line pairs are amplified by the sense amplifier SA. Subsequently, an input circuit starting signal .phi..sub.WRA generated as a result of the transition of the control input WE to the low potential is transited to the high potential. As a result, the data given to the write data DI are transmitted as the difference between signals to write line pairs IA and IA. Moreover, the write Y-select signal YWA1 is transited to the high potential to turn on the write switch WS connected with the write destination memory cell, As a result, the write line pairs IA and IA are conducted to the data line pairs DA1 and DA1, respectively. As a result, the data line pairs DA1 and DA1 are set to the potential corresponding to the data fed to the write data DI. After this, the input circuit starting signal .phi..sub.WRA is transited to the low potential, but the potential of the data line pairs is held by the sense amplifier SA. In the data line pairs for which the write switch WS is not turned on, the signal read out at first is amplified as it is by the sense amplifier so that the rewrite is executed. When the chip select signal CSA takes the high potential after the end of the rewriting operation, the selected word line WA1 and the write Y-selection signal YWA1 are shifted to the low potential. After this, the signals PPA and PNA are transited to Vcc/2 to turn off the sense amplifier SA and to transit the precharge signal .phi..sub.PA to the high potential. As a result, the data line pairs are shorted and have their potential set to the precharge potential VH so that the initial state is restored. The operations described above are the writing operations.
Here, the foregoing description is directed to the case in which a memory cell in the array A is continuously subjected to the reading operation and the writing operation. Despite this description, however, one of the reading and writing operations can be continuously executed. Moreover, it is quite natural that memory cells in desired positions in a memory cell array, which are different for every reading operation or writing operation, can be subjected to the reading operation or the writing operation by switching the modes AR, AW, BR and BW.
Next, the operations of the arithmetic mode will be described. FIG. 13(b) shows the operation waveforms for obtaining a neuron output value V.sub.12. Assume that the necessary connection weights and neuron output values or the like have already been written by the writing operations in the memory mode. In order to establish the mode NR, first, the chip select signals CSA and CSB are set to the low level, and the write control signal WE is set to the high level whereas the arithmetic circuit control signal NE is set to the low level. The addresses ADDA and ADDB are so set as to select the word line WA1 of the array A and the word line WB1 of the array B. Since the signals CSA and CSB are at the low level, the precharge signals .phi..sub.PA and .phi..sub.PB are transited to the low level. Since the signal NE is at the low level, the arithmetic circuit starting signal .phi..sub.N is transited to the high level. Subsequently, the word lines WA1 and WB1 are selected so that the neuron output values V.sub.11, V.sub.21, - - - , and V.sub.n1 and the connection weights T.sup.1.sub.11, T.sup.1.sub.12, - - - , and T.sup.1.sub.1n, are read out of the memory cells through the word line WA1 onto the data lines. Thus, the neuron output values read out from the array A and the connection weights read out from the array B are input to the multiplier MT which has been started by the arithmetic circuit starting signal .phi..sub.N, as shown in FIG. 11. In the multiplier MT, the data lines at the side of array A and the data lines at the side of array B are individually connected with the gates of the MOS transistors (T.sub.1, T.sub.2), which in turn are connected with the product/sum output line NO and the dummy line DM through the switching MOS transistor (T.sub.3) fed with the arithmetic circuit starting signal .phi..sub.N. The product/sum output line NO has its one terminal connected with the power source VM through the load RM1, and the dummy line DM has its one terminal grounded to earth. When the signal read out to the data lines is amplified to Vcc or 0 V by the sense amplifier SA, the electric current flows from the power source VM, through the load RM1, and to the earthed electrode in the multiplier, in which the product of the neuron output value and the connection weight is 1. As a result, the potential of the product/sum output line NO drops in proportion to the number of combinations, in which the product of the neuron output value and the connection weight is 1. The product/sum NO is input to the nonlinear transfer function circuit D. In this nonlinear transfer function circuit D, the summation of the products of the neuron output values and the connection weights is so high that the detection result of whether or not the potential of the product/sum output line NO is lower than reference voltage VR is output to the line NV. In the waveforms of the product/sum output line NO shown in FIG. 13(b), solid lines show the case in which the result of the product sum is small while the broken curves show the case in which the result of the product sum is large. The input/output circuit detects the result of the nonlinear transfer function circuit D and outputs the neuron output value V.sub.12, which is to be subsequently written in the memory cells, to the write line pairs IA and IA. FIG. 13(b) shows the waveforms of the write line pair IA. This pair IA takes the high level, as indicated by a broken curve, in the case where the product sum is large, and the level is low, as indicated by a solid curve. In the case where the product sum is small, the word line pair IA takes an opposite phase. At the time when the neuron output values are output to the write line pairs IA and IA, a latch signal .phi..sub.L is transited to high potential. As a result, the potentials output to the write line pairs IA and IA are latched by a latch circuit which is disposed in the input/output circuit IO. The latch signal .phi..sub.L may be delayed until the signals appear in the pairs IA and IA, in response to the fall of the arithmetic circuit starting signal NE. Subsequently, the arithmetic circuit starting signal .phi..sub.N is transited to the low potential to turn off the arithmetic circuit so that the data lines are precharged as in the memory mode after the word lines have fallen. At this time, the latch signal .phi..sub.L is left at the high potential so that the neuron output values output to the write line pairs IA and IA are held constant.
Next, the mode shifts to the NW mode, i.e., the later half of the arithmetic mode. First, the chip select signal CSA and the write control signal WE are set to the low level, and the chip select signal CSB is set to the high level in order to switch the address ADDA so that the memory cell for writing the neuron output value in the array A may be selected. The arithmetic circuit starting signal NE is left at the low level. As a result of the fall of the signal CSA, the precharge signal .phi..sub.PA drops to the low level to establish a state in which array A may be written. Subsequently, the potentials of the selected word line WA2 and the write Y-selection signal YWA1 fall. As a result, the neuron output value V.sub.12, output to the write line pairs IA and IA, is written in the memory cell which is connected with WA2 and the data line DA1. Finally, all the potentials of the word lines are dropped for the precharge. Since, moreover, the arithmetic circuit control signal NE falls, the latch signal .phi..sub.L falls to release the latch. Thus, preparations are made for the next operations. The operations described thus far are all in the arithmetic mode. All the neuron output values can be calculated by continuing similar operations in different addresses in accordance with FIG. 8(d).
In the structure thus far described, the circuit of the multiplier MT, which is connected with the dummy line DM, may be omitted. If, however, the gate capacitors, or the like, in the MOS transistors of the multiplier MT are added to the data lines only at one side, the data line capacities are unbalanced which troubles the operations of the sense amplifier, whatever the case may be. Using this scenario, the inferior influences coming from the unbalance of the data line capacities could be avoided in FIG. 11.
Next, an embodiment of the circuit suitable for use in FIG. 11 is shown. FIG. 14(a) shows one embodiment of the nonlinear transfer function circuit D. The present embodiment is constructed of: a differential amplifier composed of bipolar transistors Q720 and Q719, a resistor R72, and a MOS transistor Q721; and an inverter composed of an inverter INV75, MOS transistors Q715, Q716, Q717 and Q718, a resistor R71 and a diode D71. The present circuit is initially started when the signal .phi..sub.N takes the high potential. FIG. 14(b) shows a relation between the potential of the product/sum output line NO, the input of the nonlinear transfer function circuit D, and the potential of the output NV. The output NV takes the high potential, if the potential of the product/sum output line NO is lower than the reference potential VR, but takes the low potential if the potential of the line NO is higher than the reference potential VR. Since, according to the present embodiment, bipolar transistors are used in the differential amplifier, it is possible to realize a nonlinear circuit which is characterized to have a steep rise for the change in the input. BY setting the reference potential VR to a desired value, moreover, the characteristics of the nonlinear transfer function circuit D can be easily changed. Incidentally, the output of the differential amplifier cannot be made so high so as to avoid the saturation of the bipolar transistor Q719. As a result, the downstream inverter may not operate if it is connected directly with the output of the differential amplifier. Therefore, the resistor R71 and the diode D71 are provided to adjust the potential, which is to be input to the MOS transistor Q717, lower.
FIG. 14(c) shows one embodiment of the input/output circuit IO. The write circuit WR is composed, as shown in FIG. 14(c), of an input buffer INBUF, write switches SWA and SWB, a latch circuit LAT, and inverters INVIA and INVIB. The write switches SWA and SWB are used to switch whichever array, A or B, has its memory cells written with the write data DI. When the switching signal .phi..sub.WRA is at the high potential, the write data DI is written through the input buffer INBUF to the memory cells of array A by the write line pairs IA and IA. When the switching signal .phi..sub.WRB is at the high potential, the write data DI is written through the input buffer INBUF to the memory cells of array B by the write line pairs IB and TB. The latch circuit LAT latches the data, which are output to the output NV of the nonlinear transfer function circuit D in the arithmetic mode, to write them in the memory cells of array A by the write line pairs IA and IA. Since the potential relationship of the output NV of the nonlinear transfer function circuit D and the write line pairs IA and IA is in phase, which is apparent from FIG. 14(c), the potential relationship between the product/sum output line NO of the nonlinear transfer function circuit D and the input common line IA is opposite one another, as shown in FIG. 14(d). Since, in the embodiment of FIG. 11, the potential of the product/sum output line NO is the lower of the larger product sum of the neuron output values and the connection weights, as has been described hereinbefore, the circuit is made such that the potential relationship between the product/sum output line NO and the input common line IA be opposite to one another. Since in this case the potential of the product/sum output line NO is designed so that it may increase more for larger product sums of the neuron output values and the connection weights, it is quite natural that the circuit be made so that the potential relationship between the product/sum output line NO and the input common line IA be in phase.
FIG. 14(e) shows one embodiment of the read circuit OUT. This read circuit OUT is composed of a current/voltage converter IVOUT1, a level shift circuit LS, a read latch circuit OUTLT and an output buffer BUFOUT. In the current/voltage converter IVOUT1, the data read out as the impedance differences to the read lines OA and OA and the read lines OB and OB are converted to the differences of the voltages of the lines OA' and OA' and read lines OB' and OB'. In the level shifts LS, the voltage of the data read out from the current/voltage converter IVOUT1 is shifted to a level, at which the bipolar transistors in the downstream read latch circuit OUTLT are not saturated, to transmit it to the read latch circuit OUTLT.
A detailed embodiment of the read latch circuit OUTLT is shown in FIG. 14(f). The read differential amplifiers AMPA and AMPB in the read latch circuit OUTLT are used to switch the data read out from the memory cells of array A through the read lines OA and OA to lines L1 and L2 with the data read out from the memory cells of array B through the read lines OB and OB to lines L3 and L4 and are to be read out as the read data DO. When a switch signal .phi..sub.A is at the high potential, the data read out from the memory cells of array A are output as the read data DO. When a switch signal .phi..sub.B is at the high level, th data read out from the memory cells of array B are output as the read data DO. In the read latch circuit OUTLT, the bipolar transistor Q1A is turned off, but the bipolar transistor Q1B is turned on when the read latch circuit .phi..sub.LR is transited to a higher potential than the voltage V.sub.B2. As a result, the differential amplifiers AMPA and AMPB are turned off, and the differential amplifier AMPC is turned on. As a result, the read data are latched by the differential amplifier AMPC and the level shift circuit LSC. Specifically, according to the present embodiment, the read data DO can be latched and output continuously for a desired period of time by transmitting the read latch circuit .phi..sub.LR to a higher potential than the voltage V.sub.B2 after the read data have been fixed.
Incidentally, in the embodiment of the multiplier MT shown in FIG. 11, the data line pairs of array A are connected with the gates of the MOS transistors which are farther from the earthed electrode than the data line pairs of array B. As a result, when a product is to be calculated, the neuron output values and the connection weights are not handled equivalently. If this raises a problem, the embodiment in FIG. 15 may be used. In FIG. 15, the data line DAi is connected with the gates of MOS transistors Q7C3 and Q7C6, and the paired data line DBi is connected with the gates of MOS transistors Q7C5 and Q7C4. Since the two data lines are connected with the MOS transistors closer to the earthed electrode and the MOS transistors farther from the same, the neuron output values and the connection weights are handled equivalently handled. As has been described hereinbefore, according to the embodiment shown in FIG. 11, the embodiment shown in FIG. 8(a) can be realized by using the DRAM cell which is composed of one MOS transistor and one capacitor. The DRAM cell can have its occupied area drastically reduced in order to attain the benefit that it be highly integrated over the chip. Although not discussed earlier, in the DRAM cell composed of one transistor and one capacitor, refreshing operations are necessary within a constant time period for compensating the reduction of the stored charges due to the current leakage of the capacitor. In the present invention, too, the refreshing operations can be easily accomplished, if necessary, like the ordinary DRAM no matter which mode might be used, the memory mode or the arithmetic mode.
In the above-mentioned embodiment, the memory cells are exemplified by the DRAM cells but should not be limited thereto. A similar data processing system can also be realized by using even other memory cells. Next, an embodiment using SRAM cells will be described. FIGS. 16(a) and 16(b) are circuit diagrams showing SRAM cells MCS. In the embodiment of the present invention, the SRAM cells MCS are used as the MC of FIGS. 7(a) and 7(b), FIG. 8(a), FIG. 9(a) and FIG. 10(a). Either of the embodiments, shown in FIGS. 16(a) and 16(b), is advantageous in that the controls are far easier than the case which uses DRAM cells because it requires neither a rewriting nor refreshing operation. FIG. 16(c) shows one embodiment for realizing the embodiment of FIG. 8(a) by using the SRAM cell of FIG. 16(a) or 16(b). In FIG. 16(c), letters MCS designate the SRAM cell, and letters LD designate a data line load. FIGS. 17(a) and 17(b) show examples of the operation waveforms. FIG. 17(a) shows an example of the case in which cells connected with the data lines DA1 and DA1 and the word line WA1 are continuously subjected to the reading and writing operations in the memory mode. FIG. 17(b) shows an example of the operation waveforms of the case in which the neuron output valve V.sub.12 is to be calculated in the arithmetic mode from both the neuron output values V.sub.11, V.sub.21, - - - , and V.sub.n1 which are stored in the memory cells of the word line WA1, and the connection weights T.sup.1.sub.11, T.sup.1.sub.21, - - - , and T.sup.1.sub.1n which are stored in the memory cells of the word line WB1. The basic operations are similar to those of the aforementioned case of the DRAM cells, and their description will be omitted. The SRAM cell has the benefit that its control is simpler than the DRAM cell, because it does not need the rewriting operation and the refreshing operation. Because there is no necessity for the rewriting operation, moreover, another benefit is that the reading and writing speeds in the memory mode and the cycles in the arithmetic mode can be accelerated.
The description thus far made is directed to the example of the circuit structure, in which the embodiment of FIG. 8 is realized by using the DRAM cells and the SRAM cells. Next, an example of the circuit structure for expressing the neuron output values and the connection weights by using a plurality of memory cells will be described. Although the embodiment to be described uses DRAM cells, the present invention can likewise be realized even using SRAM cells.
Next, an example of the circuit structure in which the neuron output values are expressed by using the DRAM cells and in which the connection weights are expressed by using a plurality of memory cells will be described. In FIG. 18(a), the data line pairs DA11 and DA11, DA12 and DA12, - - - , and DALP and DALP in array A correspond to the data line pairs in array A, which are to be input to the adder al in FIG. 9(a). On the other hand, the data line pairs DAn1 and DAn1, and DAn2 and DAn2, - - - , and DAnP and 8DAnP in array A correspond to the data line pairs in array A, which are to be input to the adder an in FIG. 9(a). Array B has similar correspondences. As shown in the input/output circuit DI010, r number of input terminals DO1, - - - , and DOr and r number of output terminals Di1, - - - , and DIr are provided (wherein r is the larger number of p and q) so that data made up of p bits or q bits indicating the neuron output values or the connection weights may be simultaneously read out or written in memory mode. In array A in the arithmetic mode, the data of every p bits read out to the data lines by selecting the word lines are synthesized by the adder ADD to output the neuron output values to the neuron output values to the neuron output value output lines VO1, VO2, - - - , and VOn. In array B, on the other hand, data made up of q bits read out to the data lines by selecting the word lines are synthesized by the adder ADD to output the connection weights to connection weight output lines TOl, TO2, - - - , and TOn. These values are input to the BLK2 so that the resultant product sum is input to the nonlinear transfer function circuit D10. The output of the nonlinear transfer function circuit A10 corresponding to the neuron output value is transmitted to the input/output circuit DOO10 and latched by the latch signal .phi..sub.L. Subsequently, the address is switched to select the p number of cells to write the determined neuron output value, and the write Y-select signal YWAi is raised to write the latched neuron output values in parallel in the p number of selected cells. By continuing these operations, the neuron output values can be updated like in the embodiment of FIG. 11(a). According to the present embodiment, the embodiment of FIG. 9(a) can be realized by equivalently adding the data of the plurality memory cells input to the adder ADD. By weighing and adding, bit by bit, the data of the plurality memory cells input to the adder ADD, moreover, it is possible to realize the embodiment of FIG. 10(a), in which the neuron output values and the connection weights are expressed with binary numbers of a plurality of bits. Since the present embodiment can also be applied to the case in which the neuron output values and the connection weights are expressed with a plurality of bits by another method, a variety of data processings can be accomplished in accordance with the purpose. Since the DRAM cells are used in the present embodiment, a high integration can be achieved. Since, moreover, the data of the plurality of memory cells are processed in parallel, both in the memory mode and in the arithmetic mode, data processing can be executed at high speeds like in the case in which expressions have 1 bit, although the neuron output values and the connection weights are expressed by the plurality of bits. Here, in BLK1, the signals of the plurality of memory cells are synthesized by the adder, and the result is input to BLK2 which acts as the product/sum circuit. However, a variety of modifications can be made by omitting the addition at so that the data of the plurality of memory cells indicating the neuron output values or the connection weights are input in parallel to the product/sum circuit BLK2 so that they may be subjected to multiplications and summations.
In the following embodiment, FIG. 18(b) for realizing the embodiment of FIG. 9, in which the neuron output values and the connection weights are expressed in a plurality of equivalent bits by the embodiment shown in FIG. 18(a), shows one embodiment of BLK1 of FIG. 18(a). BLK1 is shown which is connected with data lines DA11, - - - , and DA1P of array A. The same circuit can also be used in another BLK1 of array A. The circuit of the present embodiment can also be used in array B, if the number of the data line paris, the read line pairs or the write line pairs is changed from p to q and if a q number of circuits, each having a p number of precharge circuits PR, are provided. In the present embodiment, p pairs of read line pairs OA1 and OA1, - - - , and OAp and OAp and p pairs of write line pairs IA1 and IA1, - - - , and IAp and IAp are provided so that the p number of memory cells may be subjected to the writing or reading operations in parallel. The read sense amplifier RSA and the write switch WS are consecutively connected in the same BLK1, as shown, with the read line pairs OA1 and OA1, - - - , and OAp and OAp and the p pairs of write line pairs IA1 and IA1, - - - , and IAp and IAp. For one pair of read or write lines, specifically each p pair is connected with the data line pair. The adder ADD is composed of a load circuit LD103 and a p number of voltage/current converters VI. In the voltage/current converter VI, the data lines DA11, DA12, - - - , and DA1p are connected with the gates of the MOS transistors, which in turn are connected in series with the MOS transistors having their gates fed with the arithmetic circuit starting signal .phi..sub.N, to connect the earthed electrode and the neuron output value output line VO1. This neuron output value output line VO1 is connected through a resistor in the load circuit with the power source VMO1. As a result, if the amplification of the data line potential ends in the state started by the arithmetic circuit starting signal .phi..sub.N, the potential of the neuron output value output line VO1 is dropped by a voltage proportional to the number of data lines which are raised to the high potential, i.e., Vcc. According to the present embodiment, therefore, the neuron output values can be expressed in terms of the potential drop of the neuron output value output line VO1. Incidentally, the provision of similar circuits at one side of the data lines DA11, - - - , and DA1p is to avoid the unbalance of the data line capacities for the same reasoning as that of the multiplier MT of FIG. 11(a). According to the embodiment thus far described, the neuron output values or the connection weights expressed by the plurality of memory cells can be read out to the neuron output value output line or the connection weight output line.
FIG. 18(c) shows one embodiment of the block BLK2 for calculating the product sum of the neuron output values and the connection weights and the nonlinear transfer function circuit D10. In FIG. 18(c), block BLK2 is composed of the load circuit LD102 and the multiplier MT10. The neuron output value output lines VO1, VO2, - - - , and VOn and the connection weight output lines TO1, TO2, - - - , and TOn are connected with the gates of the MOS transistors M16c1 and M16c2 of the MT10, and the MOS transistors are connected in parallel with the MOS transistors M16c3, which has its gate fed with the arithmetic circuit starting signal .phi..sub.N, to connect the earthed electrode and the product/sum output line NO. On the other hand, the product/sum output line NO is connected through the resistor RO2 in the load circuit LD102 with the power source VMO2. In the state in which the arithmetic circuit starting signal .phi..sub.N is high, so that the present circuit is started, the potential of the product/sum output line NO is dropped for the larger sum of the products of the potentials of the corresponding neuron output value output lines VO1, VO2, - - , and VOn and the connection weight output lines TO1, TO2, , and TOn. As has been described hereinbefore, the potentials of the neuron output value output lines VO1, VO2, - - - , and VOn and the connection weight output lines TO1, TO2 and TOn are dropped substantially in proportion to the magnitudes of the neuron output values and the connection weights so that the potential of the product/sum output line NO becomes higher for the larger product sum of the neuron output values and the connection weights. The product/sum output line NO is input to the nonlinear transfer function circuit D10. The nonlinear transfer function circuit D10 can be constructed by connecting n number of circuits shown in FIG. 18(d) in parallel. The circuit of FIG. 18(d) is made like the nonlinear transfer function circuit D of FIG. 14(a) by combining the differential amplifier and the inverter. Since, however, the polarities of the product/sum output line NON and the product sum of the neuron output values and the connection weights are different between the embodiments of FIG. 11 and FIGS. 18(a), 18(b) and 18(c), the resistor Rx of the differential amplifier of FIG. 18(d) is connected in the opposite position to resistor R72 of FIG. 14(a). In FIG. 18(d), therefore, the output NVx turns to the high potential if the product/sum output line NO exceeds the reference voltage VRx (x=1, 2, - - - , and p). If p number of such nonlinear transfer function circuits DSx are provided and if the reference voltage VRx is changed, as shown in FIG. 18(e), the change in the product/sum output line NO can be indicated by the number of the p outputs NVx that take the high potential. According to the present invention, the characteristics of the nonlinear transfer function circuit can be easily varied by varying the value of the reference voltage VRx. Incidentally, in the case which the circuit shown in FIG. 18(c) is used as the multiplier MTO10, the potential variation of the product/sum output line NO is generally kept away from linearity for the magnitudes of the product sum of the neuron output values and the connection weights by the characteristics of the MOS transistors. It is, therefore, advisable to set the value of the reference voltage VRx by considering the characteristics of the multiplier or the adder so that the characteristics of the nonlinear transfer function circuit may take a desired shape. As the case may be, the characteristics of the individual chips may be made difficult to understand accurately because of the fluctuations of the production conditions. In this case, the known neuron output values and connection weights are actually written in arrays A and B, and the potential of the product/sum output line NO in the arithmetic mode is measured so that the value of the reference voltage VRx may be resultantly trimmed to the desired characteristics.
Incidentally, the details of the input/output circuit DIO10 of FIG. 18(a) will be omitted here. The circuit for reading or writing a plurality of memory cells in parallel can be easily realized by using a plurality of circuits which are similar to the read circuit OUT or the write circuit WR shown in FIGS. 14(c), 14(e) and 14(f). Moreover, the structure of the clock generator 16 will be omitted but can be easily realized like the circuit used in the ordinary memory.
Next, the method for realizing the embodiment of FIG. 10, in which the neuron output values and the connection weights are binarily expressed in a plurality of bits, will be described in connection with the embodiment of FIG. 18(a). In order to add the data expressed binarily with a plurality of bits, as shown in FIG. 10(a), it is necessary to weigh and add the data of the plurality of memory cells bit by bit. For this to take place, if the ratios of the gate width of the MOS transistors connected with the data lines in the voltage/current converters VI1, VI2, - - - , and VIp in FIG. 18(b) are 1:2:4:, - - - ,and :2.sup.P, the potential of the neuron output value output line VO1 drops in proportion to the magnitude of the binary neuron output values. If, therefore, similar circuits are used for other neuron output values or connection weights, the weighing additions can be realized, as shown in FIG. 10(a). The block BLK2 shown in FIG. 18(c) can be used because it is the multiplier. The nonlinear transfer function circuit has to be given the function of the AD converter of rewriting the arithmetic result output to the product/sum output line NO in the plurality of memory cells in binary notations. For this to take place, it is possible to use the embodiment shown in FIG. 19(a). In the embodiment of FIG. 19(a), a z (z=2.sup.P) number of nonlinear transfer function circuits DS1, DS2, - - - , and DSz and an encoder are combined. These nonlinear transfer function circuits DS1, DS2, - - - , and DSz are given the characteristics shown in FIG. 19(b) by adjusting the reference voltage VRx with the circuit of FIG. 18(d). Then, the magnitude of the product sum of the neuron output values and the connection weights can be known like the embodiment of FIG. 18(c) from the number of outputs NA1, NA2, - - - , and NAz that have a high potential. Then, the equivalent expressions of z bits have to be changed into binary expressions of p bits by the encoder since they have to be transmitted to the write circuit through the p number of output lines NV1, NV2, - - - , and NVp. It follows that the encoder of FIG. 19(a) may be given the input/output relationship shown in FIG. 19(c). This encoder can be realized without difficulty. An example of the structure for p=3 is shown in FIG. 19(d). The present embodiment can be easily extended to the cases other than that for p=3.
The description made thus far is exemplified by the multi-layered neural network. Despite this exemplification, however, the present invention should not be limited to the multi-layered neural network but can be applied to other types of networks by using teh embodiments thus far described. FIGS. 20(a) and 20(b) and FIGS. 21(a) and 21(b) show embodiments for realizing the data processing using the Hopfield network according to the algorithm of FIG. 5(b). FIG. 20(a) shows an embodiment in which the unsynchronized Hopfield network is realized by using memory cells one by one for expressing the neuron output values and the connection weights. As have been described with reference to FIGS. 2 and 3, the basic arithmetic method is commonly shared between the multi-layered network and the Hopfield network. In the Hopfield network, however, the calculations are carried out by using the neuron output values from all the neurons including those of itself. In FIG. 20(a), therefore, all the neuron output values are stored in one word line of array A. In array B, as shown, the connection weights necessary for calculating one neuron output value are stored on a common word line The updating of the neuron output values can be executed in the following manner. In order to update the neuron output value V.sub.1, for example, the word line WA of array A and the word line of j=1 of array B are raised. As a result, the new neuron output value of g(T.sub.11 V.sub.1 +T.sub.12 V.sub.2 + - - - +T.sub.1n V.sub.n) is calculated. This value may be written in the memory cell which is located in the position of i=1 on the word line SWA of array A. The other neuron output values are updated similarly. The value V.sub.4, for example, is updated by raising the word line WA of array a and the word line of j=4 of array B. As a result, the new value V.sub.4 of G(T.sub.41 V.sub.1 +T.sub.42 V.sub.2 + - - - +T.sub.4n V.sub.n) is calculated. This value may be written in the memory cell in the position of i=4 on the word line WA of array A. Thus, the calculations of the unsynchronized Hopfield network can be executed by updating the neuron output values V.sub.i in the desired order. The calculations of the synchronized Hopfield network can be easily realized by using the memory cells on the word line WA1 of array A for storing the neuron output values at present and by using the memory cells on the word line WA2 for storing the new neuron output values, as shown in FIG. 20(b). First, the word line WA1 of array A and the word line of j=1 of array B are raised. As a result, the new value V.sub.1 of g(T.sub.11 V.sub.1 +T.sub.12 V.sub.2 + - - - +T.sub.1n V.sub.n) is calculated. This value may be written in the memory cell in the position of i=1 on the word line WA2 of the array A. Subsequently, the neuron output values V.sub.2, V.sub.3, - - - , and V.sub.n are updated and written in the memory cells on the word line WA2 of array A. When all the neuron output values have been updated, the process is continued by interchanging the roles of word lines WA1 and WA2 of array A such that word line WA2 is selected for calculating the neuron output values and word line WA1 is selected for storing the neuron output values. From now on, the processes by interchanging the roles of word lines WA1 and WA2 of array A. Thus, the calculations of the synchronized Hopfield network can be executed according to the embodiment of FIG. 20(b).
Likewise, the Hopfield network can be realized by using a plurality of memory cells for expressing the neuron output values and the connection weights. FIG. 21(a) shows an embodiment for realizing the unsynchronized Hopfield network by using p and q number of memory cells equivalently for expressing the neuron output values and the connection weights. Like FIG. 20(a), all the neuron output values are stored in one word line of array A. Here, p number of cells express one neuron output value Array B is stored such that the connection weights necessary for calculating one neuron output value are arrayed on a common word line for every q number of cells. The neuron output values may be updated as in the embodiment of FIG. 20(a). Since, however, p number of memory cells are used individually for expressing the neuron output values, p number of output lines of nonlinear transfer function circuit D are provided so that the calculation results may be written in parallel in p number of cells. The synchronized Hopfield network can also be easily realized like FIG. 21(b) if two word lines of array A are used as in FIG. 20(b). Likewise, it is quite natural that the synchronized and unsynchronized Hopfield networks can be realized by binary expressions using p and q number of memory cells for expressing the neuron output values and the connection weights, as shown in FIG. 10(a).
FIGS. 8(a) and FIGS. 20(a) and 20(b), and FIG. 9(a) and FIGS. 21(a) and 21(b) present basically identical structures. If, therefore, the embodiments of FIGS. 11 to 19 are used, the data processings according to the embodiments of FIGS. 20(a) and 20(b) and FIGS. 21(a) and 21(b) can be easily realized. Incidentally, in the Hopfield network, the procedure of continuing the updating of the neuron output values falls in the so-called "local minimum", in which the energy is not the minimum, but minimal, so that the neuron output values are no longer varied. In order to avoid this, the well-known "quasi-annealing method" can be used. The method of changing the shape of the nonlinear transfer function gradually is known for realizing the quasi-annealing method, as is described on pp. 122 of Neural Network Processing (published by Sangyo Tosho and edited by Hideki Asou). According to the present invention, this method can be easily realized by switching a plurality of nonlinear transfer function circuits D having different characteristics and by controlling the characteristics of the nonlinear transfer function circuits D externally.
Although examples in which the neuron output values and the connection weights have been handled as positive number mainly in the multi-layered or Hopfield network have been described, it may be convenient depending upon the application that either both or one of the two values can take positive and negative values. The present invention can be easily applied to such a case. FIG. 22 shows one embodiment of the present invention, in which both the neuron output values and the connection weights are enabled to take positive and negative values. In FIG. 22, the neuron output values are stored in memory cell array A, and the connection weights are stored in memory cell array B. The individual values are expressed with p or q bits, indicating the absolute magnitudes, and with 1 bit indicating the codes. The bits indicating the codes (as will be called the "coding bits") indicate a positive value with "1" and a negative value with "0". Of the neuron output values and connection weights thus read out by the method similar to those described earlier, the portions of the p or q bits indicating the absolute values are input to the adders a1, - - - , and an and b1, - - - , and bn so that the resultant analog values are input to the multipliers m1, - - - , and mn. Incidentally, when the neuron output values and the connection weights are to be expressed in binary notation, the data of the individual p and q bits input to the aforementioned adders a1, - - - , and an, and b1 - - - , and bn may be weighed and input like FIG. 10. On the other hand, the coding bits are input, as shown in FIG. 22, to exclusive OR circuits EOR1, - - - , and EORn. In the case in which the coding bits fail to become similar, namely, when the result of a multiplication is negative, the outputs of the aforementioned exclusive OR circuits become high. In case of similarity, namely, when the multiplication result is positive, the outputs of the exclusive OR circuits become low. Switches SW1, - - - , and SWn operate to transfer the outputs of the multipliers to adders C1, - - - , and cn, when the outputs of the exclusive OR circuits become low, or to the multipliers c1', - - - , and cn' when the outputs of the exclusive OR circuits become high. As a result, the sum of the positive results of multiplications is output to the product/sum output line NO, and the sum of the negative multiplication results is output to the product/sum output line NO'. In the nonlinear transfer function circuit D, the difference between the signals of the product/sum output line NO and the product/sum output line NO' is converted into a digital value of p bits and fed to the bus ABS so that the coding bits are determined according to the magnitudes of the signals of the product/sum output line NO and the product/sum output line NO' and output to a bus SIGN. Incidentally, it is quite possible, according to methods similar to the ones mentioned above, to give the nonlinear characteristics, as shown in FIG. 9 or 10, to the expressions of the neuron output values. According to the present embodiment, both the neuron output values and the connection weights can take positive and negative values. As a result, the present embodiment is advantageous in that the range for applying data processing is extended. Although both the neuron output values and the connection weights are able to take positive and negative values, it is easily possible to make modifications so that either one takes positive value.
The description thus far made has been directed to the embodiment in which the product/sum function necessary for calculating the neuron output values and the nonlinear transfer function circuit are realized in the arithmetic circuit. Despite this description, however, a circuit for other calculations can be added to the arithmetic circuit. For example, the data processing system according to the present invention can be applied to the so-called "classification problem" such as the speech recognition or the letter recognition, in which input patterns are classified into several classes. In this case, the comparator is conveniently disposed in the arithmetic circuit, as has been described earlier. In the classification problem, a desired value corresponding to a certain class can be attained as the output in the case where the input patterns are apparently classified into the class. In the delicate case, however, in which it is questionable to determined which of the plurality of classes the input pattern belongs to, the classification may fall in the middle, between the desired values of the plurality of classes. In the case where the input speech is `K` in the speech recognition, for example, it is coded, and the connection weight is set so that the neuron output value (or the desired value) of 1111 may be obtained in the output layer of the speech waveforms in the input layer. If the input is `C` the connection weight is set so that the output value (or the desired value) of 0000 may result. Ifs in this case, a middle speech waveform between `K` and `C` results, the neuron output value of the output layer may give a middle value such as 0001 or 1110. In this case, the distance (or similarity) between the neuron output value of the output layer and the desired value of 1111 of the `K` or the desired value 0000 for the `C` can be interpreted as measures of the similarity to the `K` or `C` of the input speech. It is, therefore, convenient to give a function to determine the distance between the output result and the desired value by providing the arithmetic circuit with a circuit for comparing the neuron output value of the output layer and the desired value of the class.
FIG. 23 shows one embodiment which integrates in one semiconductor chip: an arithmetic circuit 12a for comparing the neuron output values and the desired value; and an arithmetic circuit 12b for calculating the neuron output values. In FIG. 23: the desired value is stored in the memory TG; the neuron output values are stored in memory A; and the connection weights are stored in memory B. The calculations of the neuron output values may be realized by the method similar to those thus far described, by reading the neuron output values from memory A and the connection weights from memory B, by calculating the neuron output values with the arithmetic circuit 12b and by writing the calculated result in memory A. The comparisons are carried bout by reading the neuron output values from memory A and the desired value from memory TG, by determining the distances in parallel with the arithmetic circuit 12B, and by writing the result in memory TG or sending the same through the input/output device. Since, in the present embodiment, both memories TG and A and the arithmetic circuit 12a are formed on the common chip, the number of buses 1 and 2 can easily be increased to process the numerous bits in parallel. This results in the benefit that the distances can be calculated at high speed. Incidentally, in the structure so far described, it is convenient to divide the arithmetic mode into a mode called the neuron output value calculating mode which calculates the neuron output values and a mode called the comparison mode which compares the neuron output values and the desired values to determine the distance. The switching of the arithmetic mode, for example, can be accomplished in response to the two arithmetic circuit control signals NE1 and NE2. Specifically: the memory mode may be selected if both the signals NE1 and NE2 are high; the neuron output value calculating mode may be selected if the signal NE1 is low whereas the signal NE2 is high; and the comparison mode may be selected if the signal NE1 is high whereas the signal NE2 is low. Incidentally, in the embodiment of FIG. 23, the memory is divided into thirds whereas the arithmetic circuit is divided into halves, but these divided circuits may naturally be mixed on the chip. As has been described above, according to the present embodiment, it is possible to determine at high speed, the distances between the neuron output values and the desired value. As a result, the data processing speed can be accelerated when necessary, as in pattern recognition which uses the multi-layered network to compare the neuron output values and each desired value to determine the distances between them.
FIG. 24 shows one embodiment of the arithmetic circuit 12a in FIG. 23, i.e., a circuit for comparing the neuron output values of the output layer and the desired value to calculate the hamming distances inbetween. In the following, it is assumed that memories TG and A of FIG. 23 are of the type, in which the data of the memory cells are read out to the data line pairs, as in FIG. 11, FIG. 16 or FIG. 18, and that the memories have arrays RG and A, respectively. The circuit of FIG. 24 is composed of a comparator CMP and a comparison result converter COMPOUT. The comparator CMP is composed of a comparator CMPU and a load resistor RCMP connected in parallel, and the comparison result converter COMPOUT is composed of differential amplifiers AMP211, AMP212, - - - , and AMP21Z. The comparator CMP is connected with the data lines DTG1 and DTG1, - - - , and DTGr and DTGr of the array TG, and the data lines DA1 and DA1, - - - , and DAr and DAr of array A. Here, letter r designates the number of memory cells on one word line and takes the value of n, if the neuron output values are expressed in 1 bit, and the product of n and p if the neuron output values are expressed in p bits. According to the present embodiment, it is possible to calculate the hamming distance between the data read out onto the data lines DTG1 and DTG1, - - - , and DTGr and DTGr of array TG and the data read out onto data lines DA1 and DA1, - - - , and DAr and DAr of array A. The operations of the present embodiment will be described below. First, a clear signal .phi..sub.C is raised in advance to turn on a MOS transistor Q216 and to break the gate voltage of a MOS transistor Q215. After the clear signal .phi..sub.C has been broken so that the signal is read out to the data line to set the data line potential to Vcc or 0 V, the comparator is started by a comparator starting signal .phi..sub.CMP. Then, the exclusive OR is taken in each of the groups of the data lines (DTG1, DA1), (DTG2, DA2), - - - , and (DTGr, DAr) connected with the comparator. As a result, in the case where the data are similar on the data lines in array TG and the data lines in array A, the gate of the MOS transistor Q215 is left low, otherwise it is set high. Then, the MOS transistor Q215 is turned on in the comparator CMPU when the data are not the same between the data lines of array TG and the data lines of array A. As a result, for the larger number of groups of the data lines (DTG1, DA1), (DTG2, DA2), - - - , and (DTGr, DAr), in which the data are not similar, more current will flow from the power source VCMP through the load resistor RCMP to the earthed electrode. As a result, the potential of the compare line CO will be lower for larger numbers of groups in which data are not similar. The compare line CO disposed in the comparison result converter COMPOUT is connected with the differential amplifiers AMP211, AMP212, - - - , and AMP21Z. If the reference voltages VRC1, VRC2, - - - , and VRCZ of those differential amplifiers are set to suitable values, the number of comparison result output lines DCO1, DCO2, - - - , and DCOZ, which are high, is the larger drop of the potential on the compare line CO. In other words, the comparison result converter COMPOUT operates as a kind of AD converter. Thus, according to the embodiment of FIG. 24, the data read out to the plurality of data lines of array TG and the data read out to the plurality of array A can be compared to determine their hamming distance. If, therefore, one word is selected from each of the arrays TG and A, the data stored in the memory cells on the selected word lines can be compared with each other. If, therefore, the desired values are individually stored in the memory cells on array TG, it can be compared with the neuron output values which are stored in the memory cells on one word line of array A. This is done in order to know what desired value the neuron output values are close to and how close they are. Therefore, the case where the obtained neuron output values are not similar to the desired value corresponding to the class, it is possible to know at high speed what class the neuron output values are closed to and how close they are.
Incidentally, in the embodiment of FIG. 24, the result output to the comparison result output lines may be output external to the chip through the input/output circuit at each time of comparison. Alternatively, the capacity of memory TG may be made larger than that necessary for storing the desired values so that the results may be written in memory TG and then output all at once.
Finally, an embodiment for further speeding up the system of the present invention by using a register will be described. As has been described earlier, according to the present invention, the neuron output values are calculated by reading out the necessary data from the memory, by determining the neuron output values with the arithmetic circuit, and by rewriting the determined result in the memory. In other words, one arithmetic mode (i.e., the neuron output value arithmetic mode) cycle is composed of the reading operation and the writing operation, and the arithmetic circuit is inactive in the writing operation. If, therefore, the time period for which the arithmetic circuit is inactive is shortened, the speed of the arithmetic mode can be further increased. FIG. 25 shows one embodiment in which the speed of the arithmetic mode is increased on the basis of the aforementioned point of view. The embodiment of FIG. 25 is made by adding the register and the switches SW1, - - - , and SWr to the embodiment of FIG. 7. According to the embodiment of FIG. 25, the neuron output values can be calculated at high speeds by using the algorithm of FIG. 5. In the following, the description will be made upon the multi-layered network, but similar effects can be attained even in the Hopfield network. In the embodiment of FIG. 25, the output value of the first neuron of the s-th layer is calculated by raising one word line of the memory cell array A to read out the neuron output value of the (s-1)th later, by closing the switches SW1, - - - , and SWr to write the neuron output values of the (S-1')th layer in the register 14, and by opening the switches SW1, - - - , and SWr. Next, one word line of the memory cell array B is raised to read the connection weight between the neuron of the (s-1)th later and the first neuron of the s-th layer. Then the neuron output values of the (s-1)th layer are read out by register 14 so that the output value of the first neuron of the s-th layer is calculated by the arithmetic circuit 12. The calculated results are written in memory cell array A. At the same time, one word line of memory cell array B is raised to read out the connection weight between the neurons of the (s-1)th layer and the second neuron of the s-th layer, and the neuron output values of the (s-1)th layer are read out by the register 14 so that the output value of the second neuron of the s-th layer is calculated by arithmetic circuit 12. After this, the output values of the neurons of the s-th layer are likewise calculated. Next, the output values of the neurons of the (s+1)th layer are calculated by raising one word line of memory cell array A to read out the previously determined neuron output value of the s-th layer, and by closing the switches SW1, and SWr to write the neuron output values of the s-th layer in register 14, and the subsequent calculations proceed as before. As has been described hereinbefore, according to the present embodiment, the writing operation and the reading operation can be simultaneously carried out at high speed by providing register 14.
The description thus far made is directed mainly to the method of calculating the neuron output values according to the present invention, and the necessary connection weights are assumed to already have been given. Depending upon the subject, the necessary connection weights are easily given at the start or have to be determined by the so-called "learning". In the learning for the multi-layered network called "back propagation", for example, several neuron output values (or test patterns) of the input layer can be prepared in advance to determine the connection weights so that the desired neuron output values may be obtained in the output layer for the test patterns. As described in Section 2 of Neural Network Data Processing (published by Sangyo Tosho and edited by Hideki Asou), moreover, the learning algorithm for setting the connection weights such that the balanced state of the neuron output values can take the desired state even in the Hopfield network. This learning can be applied to the present invention by the following three methods. According to the first method, the learning is carried out by using an external computer, and the obtained connection weights are written in the data processing system according to the present invention. This method is advantageous since the learning can be executed by the software so that the learning algorithm can be easily changed, but it is difficult to speed up the learning. According to the second method, the arithmetic circuit of the system according to the present invention is given an arithmetic function for the learning so that the learning is executed on-chip. This method speeds up the learning but may be difficult to integrate all the circuits necessary for the learning over a common chip. The third method is an intermediate one between the first and second methods, and a portion of the calculations necessary for the learning is executed by the system of the present invention whereas the remaining portion of the calculations necessary for the learning is executed by the external computer. This method is advantageous in that it can accelerate the learning speed more than the first method and that the arithmetic circuit of the system of the present invention can be simply constructed. This third method will be specifically described below. Incidentally, the learning method is exemplified by the back propagation method in the multi-layered network. In the back propagation method (as will be shortly referred to as the "BP" method), the connection weights are updated according to the following formulas:
T.sup.s.sub.ij =T.sup.s.sub.ij +.SIGMA.d.sub.js V.sub.is -1(1);
d.sub.jm =(t.sub.j -V.sub.jm)g'(U.sub.jm) (2);
and
d.sub.js =g'(U.sub.js).SIGMA..sub.i (T.sup.s+1.sub.ij d.sub.is+1)(3)
(s=m-1, - - - , and 2),
wherein:
.SIGMA.: a small positive number;
t.sub.j : a target of the neuron output value V.sub.jm of the final layer;
g': a derivative of the nonlinear transfer function g; and
U.sub.js : a quantity before passage through the nonlinear transfer function circuit g is the j-th neuron of the s-th layer, as will be defined by the following formula:
U.sub.js =.SIGMA..sub.i (T.sup.s-1.sub.ji V.sub.is-1 +.theta..sub.js)(4).
The connection weights may be updated by determining the quantities to be updated from the above-specified formulas (1) to (4) for every input data for the learning and by using the sum of the updated quantities of all the input data for the learning. On the other hand, the updating may be carried out by adding the following term called the "inertia term" to the formula (1):
.mu..DELTA.T.sup.s.sub.ij ' (5),
wherein:
.mu.: a small positive constant; and
.DELTA.T.sup.sij : a corrected quantity of the previous updating.
The updating is continued until the difference between the neuron output values of the last layer and the target values become sufficiently small.
The learning described so far can be executed by the embodiment shown in FIG. 23 and the external computer, as will be described below. The description to be made is directed to the case in which updating is executed by summing the updated values of all the input data, but similar operations are applied to the case in which the connection weights are updated for every input data. Incidentally, the description to be made is directed to the case of the three-layered network, but similar operations are applied to the case of a network having three or more layers.
First, the input data for all the learnings and their target values are written in memories A and TG, respectively. Next, a random number having a small absolute value is written as the initial value of the connection weights in memory B. Moreover, the first input data are read out as the neuron output values of the first layer to the arithmetic circuit 12b by memory B. These values are multiplied in parallel by the aforementioned method so that the neuron output values of the second layer are calculated and written in memory A. The calculations thus far described are executed for all the learning input data to read out the neuron output values of the individual layers for the individual input data, the desired values for the individual input data, and the connection weights to the memory outside of the chip. Next, the quantities for updating the connection weights are calculated in the external computer, and the updated connection weights are written in memory B of the system according to the present invention. Incidentally, the term g' (U.sub.js) appearing in the formulas (2) and (3) may be either calculated from the value U.sub.js which is input to the nonlinear transfer function circuit D when the neuron output value V.sub.js is to calculated in the system of the present invention, or calculated inversely from the value V.sub.js by the external computer according to the following formula:
g'(U.sub.js)=g'(g.sup.-1 (V.sub.js)) (6).
In order to add the inertia term of formula (5), on the other hand, the corrected quantities of the connection weights may be stored in the memory outside of the chip for every updating so that they may be added to the newly determined corrected quantities in accordance with formula (5).
The updatings thus far described can be repeated so that learning may proceed. In order to know how the learning advances, the distances between the neuron output values of the last layer for the individual input data and their desired values can be used as measures. These distances can be calculated at high speed by using the embodiment of FIG. 24. As a result, it is easily possible to confirm the progress of learning while the learning is being accomplished.
As has been described hereinbefore, according to the present invention, the calculations of the neuron output values for the input data for the learning can be executed at high speed in the system of the present invention. In the present invention, moreover, the memories composed of memory arrays are used in memories TG, A and B so that all the input data, the desired values and the neuron output values of the preceding layer can be easily stored and so that the numerous bits can be read out in parallel by raising the word lines. As a result, the transfers of the data to the external memory can be executed altogether at high speed. As a result, the learning can proceed at high speed.
If the capacities of the memories are made sufficiently large in the present embodiment, the number of neurons can be easily changed according to the application. If, in this case, the neuron number is drastically changed, the nonlinear transfer function circuit may have to have its dynamic range changed. If this is true, it is possible to switch and use a plurality of nonlinear transfer function circuits having different characteristics and the reference voltages of the amplifiers in the nonlinear transfer function circuit. In the case in which the neuron numbers are different for the layers in the multi-layered network, the nonlinear transfer function circuit may have to have its dynamic range changed for the layers. This requirement can also be dealt with by the similar method.
Incidentally, the description thus far made is directed to the embodiment in which either the DRAM cell of the so-called "one transistor and one capacity" type and the SRAM cells shown in FIGS. 16(a) and 16(b) are mainly used, but other memory cells can naturally be used in the present invention. Since the portion for storing the connection weights need not be frequently rewritten for data processing, the kinds of cells can be changed according to the contents of the memories by using non-volatile memory cells of the DRAM cells or the SRAM cells in the portion for storing the neuron output values.
If the memory cell circuit is highly integrated by using very small memory cells such as DRAM cells of the one transistor and one capacitor, some memory cells may sometimes be inactive because the wiring lines used are very small. The neural network is advantageous in that its function is hardly influenced even if the connection weights are changed more or less, but the data processing may be troubled in the case in which the memory cells for storing the neuron output values are inactive. In order to solve this problem, the redundant word lines or data lines to be used in the ordinary highly-integrated semiconductor memory can be provided so that defective cells may not be used.
In FIGS. 14(a), 14(e) and 14(f) and FIG. 18(d), moreover, bipolar transistors are used, which can be realized by the CMOS. Still moreover, the present invention can be practiced by not only the bipolar transistors and the MOS transistors but also by other devices.
Although the foregoing description is directed mainly to the multi-layered and Hopfield networks, the present invention should not be limited thereto but can be applied to neural network data processing for networks of various types. For example, it is possible to realize the network in which the neuron output values are updated such as in the Boltzman's machine. As described on pp. 27 of Neural Network Data Processing (published by Sangyo Tosho and edited by Hideki Asou), where the Boltzman's machine is featured, although the network shape is similar to that of the Hopfield network in that the neuron output value (0 or 1) is not uniquely determined by another product sum of the neuron output values input to the neurons and the connection weights, but in a probable manner. The probability P for the neuron output value to take the value 1 is expressed by P=1/ (1+exp(-I/T)). Here, letter I designates the product sum of the no1 input to the neurons and the connection weights, and letter T designates a parameter called the temperature. The Boltzman's machine described above can be easily realized according to the present invention. For example, the reference voltage VRx of the nonlinear transfer function circuit D, as shown in FIG. 18(d), is not set to steady value but may be changed with time within the fluctuating range of the product/sum output line NO. Then, the neuron output values can be determined according to the probability. The effects obtained by changing the changing rate are similar to those obtained by changing the temperature T.
If the capacity of the memories is sufficient, as is apparent from the comparisons between FIG. 8(a) and FIGS. 20(a) and 20(b), various types of networks can be realized by a common system merely by changing the addresses of the memory cells for storing the neuron output values and the connection weights. Thus, the present invention has a very wide applicability.
Although the description thus far made is directed to the applications of neural network data processing, the present invention should not be limited thereto but can naturally realize such a system with a high degree of integration as it is used for data processing by connecting a number of processing elements having similar processing functions in the form of a network.
In the embodiments thus far described, the description is directed mainly to the structure in which the arithmetic circuit performs the analog calculations. The analog arithmetic circuit is advantageous in that it has high speed and a small circuit scale. Despite this description, however, the present invention should not be limited thereto but can be used in a digital arithmetic circuit without department from the gist thereof. In this case, the calculations can be executed very accurately by the digital arithmetic circuit.
As has been described hereinbefore, according to the present invention, the system for executing data processing by combining memories and arithmetic circuits and by performing parallel calculations with the arithmetic circuits like the parallel distributed processing system such as the neural network, in which a number of arithmetic units for relatively simple calculations are connected in network form, can be realized with a high degree of integration without sacrificing speed.
It is further understood by those in the art that the foregoing description is the preferred embodiment of the disclosed device and that various changes and modifications may be made in the invention without departing from the spirit and scope thereof.

Number	Date	Country	Kind
2-12538	Jan 1990	JPX
2-119828	May 1990	JPX

Number	Name	Date
4611299	Hori et al.	Sep 1986
4777858	Wang et al.	Oct 1988
4847755	Morrison et al.	Jul 1989
4873962	Etoh et al.	Oct 1989
4939696	Katsuo et al.	Jul 1990
4955024	Pfeiffer	Sep 1990
4974169	Engel	Nov 1990
4994982	Duranton et al.	Feb 1991
5005121	Nakada et al.	Apr 1991
5010477	Omoda et al.	Apr 1991
5023833	Baum	Jun 1991
5107442	Weideman	Apr 1992
5140670	Chua et al.	Aug 1992
5142665	Bigus	Aug 1992
5163120	Childers et al.	Nov 1992
5165010	Masuda	Nov 1992
5179680	Colwell	Jan 1993
5239654	Ing-Simmon et al.	Aug 1993
5301340	Cook	Apr 1994

	Number	Date
Parent	369163	Jan 1995
Parent	938755	Sep 1992
Parent	634046	Dec 1990

Neural network processing system using semiconductor memories

Information

Patent Number

Date Filed

Date Issued

Inventors

Original Assignees

Examiners

Agents

CPC

US Classifications

Field of Search

US

International Classifications

Abstract

Description

Claims

Priority Claims (2)

Parent Case Info

US Referenced Citations (19)

Non-Patent Literature Citations (8)

Continuations (3)

Entry
Sejnowski et al. "Parallel Networks that Learn to Prononuce English Text,"Complex Systems 1, 1987, pp. 145-168.
Advanced Micro Devices, Memory Products Data Book, Jan. 1989, pp. 4-80, 4-81.
Eberhardt et al., "Design of Parallel Hardware Neural Network Systems from Custom Analog VLSI `Building Block` Chips", International Joint Conference of Neural Networks, Jun. 18-22, 1989, pp. II-183 to II-190.
Rumelhart et al. Learning Representatives by Back Propagating Errors, Nature V. 323, Oct. 9, 1986, pp. 533-536.
Suzuki et al. "A Study of Regular Architectures for Digital Implementation of Neural Networks", 1988 IEEE International Symp. on Circuits and Systems, May 8-11, pp. 82-85.
Holler et al. "An Electrically Trainable Artificial Network (ETANN) with 10240 `Floating Gate` Synapses", Proceedings of the International Annual and Conference on Neural Network, 1989, pp. 50-55.
Manuel, "Are Artificial Neural Networks Finally Ready for Market?", Electronics, Aug. 1988, pp. 85-88.
"A 72K CMOS Channelless Gate Array with Embedded 1Mbit Dynamic RAM", Sawada et al., Proc. IEEE 1988 Custom Integrated Circuit Conference pp. 20.3.1 to 20.3.4.