Neural network circuit

A multilayer neural network circuit having a plurality of neurons each having a memory for storing a coupling coefficient, said multilayer neural network circuit further comprising:

• a control circuit which outputs the total number of non-zero coupling coefficients per layer;
• an address generator which subsequently counts the total number of non-zero coupling coefficients per layer, by the output of the control circuit;
• a plurality of characteristic data memories which store the characteristic data, and subsequently read characteristic data as the address generator counts;
• a weight memory which stores weights of neurons of the last layer;
• a plurality of coefficient memories which store coupling coefficients of the layers other than the last layer;
• a plurality of address convertors which calculate only addresses of a coefficient memory corresponding to a neuron having a non-zero coupling coefficient value between adjacent layers of the neural network circuit on the basis of the outputs of the address generator, the characteristic data memories and the control circuit, the plurality of address converters supplying the addresses thus calculated to the coefficient memories while at the same time calculating addresses to be output to the weight memory;
• a table memory which takes the coupling coefficients read out from the coefficient memories corresponding to the addresses output from the address converter and outputs total coupling coefficients obtained by multiplying the coupling coefficients of the layers corresponding to the addresses;
• a cumulative operation unit which performs cumulative additions of the product of the output of the table memory multiplied by the output of the weight memory;
• and
• an output memory which holds results produced by the cumulative operation unit at the addresses output from the address generator,
• wherein the cumulative addition is performed only on particular neurons interconnected between adjacent layers having a non-zero coupling coefficient value, and the address converter, which outputs a control signal, modifies the contents of the control signal depending upon whether an address of a non-zero coupling coefficient value and an address of a greatest non-zero coupling coefficient are in series, the multilayer neural network circuit further including a multiplexer, which receives as an input the control signal from the address converter, outputs a zero to the table memory when the aforesaid addresses are not in series, or selects and outputs the coupling coefficient from the coefficient memory to the table memory when the addresses are in series.

A neural network circuit comprising multiple layers in which:

• each said layer except a final layer comprises a plurality of quantized neurons;
• each said quantized neuron comprises one quantized signal input terminal which receives characteristic data of the respective layer and a plurality of output terminals, each said quantized neuron further comprising a coupling coefficient which changes in accordance with the value of the characteristic data coupled to said quantized signal input terminal;
• said output terminal of each quantized neuron of a given layer being coupled to each quantized neuron of a next upper layer;
• said final layer comprising a plurality of multi-input, one output neurons;
• said characteristic data input into each layer being calculated by the network having the multi-layered construction and a result is output by said final layer, whereby the given characteristic data is recognized, said neural network circuit comprising:
• address generation means, weight memory means, total coupling coefficient memory means and cumulative operation means;
• said address generation means memorizing beforehand to the respective layers the number of output terminals having a non-zero coupling coefficient value out of the plural output terminals of quantized neurons of each layer except the final layer,
• said address generation means memorizing beforehand, according to various characteristic data to be input to respective layers, an address of the first output terminal out of the output terminals having the non-zero coupling coefficient value to output an address of the first output terminal corresponding to one characteristic data to be input or calculates an address of the first output terminal having the non-zero coupling coefficient value, based on a given conversion formula which includes the characteristic data to be input and is obtained beforehand,
• said address generation means generating as an initial address a combination of addresses of the first output terminal having the non-zero coupling coefficient value of each layer corresponding to the input characteristic data, then sequentially generating combinations of addresses of the output terminals having the non-zero coupling coefficient value in each layer,
• said weight memory means memorizing beforehand each weight of the multi-input, one-output neurons of the final layer according to the combinations of various addresses which are generated by the address generation means to output a weight of one multi-input, one-output neuron corresponding to one address generated by the address generation means,
• said total coupling coefficient memory means memorizing beforehand according to the combinations of the various addresses generated by the address generation means, a total coupling coefficient which is obtained by multiplying the non-zero coupling coefficient values of each layer except the final layer to output the total coupling coefficient corresponding to one address generated by the address generation means, and
• said cumulative operation means multiplies the total coupling coefficient output from the total coupling coefficient memory means and the weight of the multi-input, one-output neuron of the final layer which is output from the weight memory means to accumulate the multiplied results.