Neural network, method of control of neural network, and processor of neural network

1. A neural network comprising:

at least one neuron core having a data input terminal for receiving analog input data, a data output terminal for outputting analog output data, and a weight value input terminal, the neuron core being configured to perform an analog multiply-accumulate operation based on the analog input data arriving from the data input terminal and a weight value arriving from the weight value input terminal, the weight value being analog data;

a weight-value-supply control unit connected to the weight value input terminal and supplying the weight value from the weight value input terminal to said neuron core,

a control-processor unit controlling the supply of the weight value from said weight-value-supply control unit and processing the analog output data output from the data output terminal at every analog multiply-accumulate operation performed by said neuron core as serial data and/or parallel data; and

a data-hold-supply control unit connected to the data input terminal and the data output terminal, and provided with at least one second register and a second selector, the second register being configured to hold at least the analog output data operated by the neuron core and output from the data output terminal such that the hold data is suppliable to output to the second selector, and the second selector being capable of selecting any one of the at least one second register or an input data arriving from a prior stage, and wherein the second selector causes opening and closing a path for transmitting the held data in the second register to the data input terminal,

wherein said control-processor unit controls timing of holding data at the second register, and timing of opening and closing of the path, and controls the selecting operation of the second selector with respect to the input data from the prior stage and the hold data in the second registers.

2. The neural network according to claim 1 , wherein said weight-value-supply control unit is provided with

a plurality of first registers respectively holding weight values and

a first selector for selecting any one of the plurality of first registers and opening and closing a path for transmitting the held weight value to the weight value input terminal.

3. The neural network according to claim 1 , wherein said control-processor unit is configured by

a main control unit controlling the overall neural network and

a sub-control unit directly controlling said neuron core, said weight-value-supply control unit, and said data-hold-supply control unit based on instructions of said main control unit,

said main control unit capable of setting the weight value at said weight-value-supply control unit and the data at said data-hold-supply control unit.

4. The neural network according to claim 3 , which can refer to the weight value set at said weight-value-supply control unit and the data set at said data-hold-supply control unit from said main control unit.

5. The neural network according to claim 1 , which compares the data set at said data-hold-supply control unit with the analog output data from the data output terminal, and notifies said main control unit of the comparison results.

6. A method of control of a neural network, the neural network comprising:

at least one neuron core having a data input terminal for receiving analog input data, a data output terminal for outputting analog output data, and a weight value input terminal, the neuron core being configured to perform an analog multiply-accumulate operation based on the analog input data arriving from the data input terminal and a weight value arriving from the weight value input terminal, the weight value being analog data,

a weight-value-supply control unit connected to the weight value input terminal and supplying the weight value from the weight value input terminal to said neuron core, and

a data-hold-supply control unit provided with at least one second register and a second selector, the second register being configured to hold the analog output data operated by the neuron core and output from the data output terminal such that the hold data is suppliable to output to the second selector, and the second selector being capable of selecting any one of the at least one second register or an input data arriving from a prior stage, and wherein the second selector causes opening and closing a path for transmitting the held data in the second register to the data input terminal,

the method comprising:

controlling the supply of a weight value from said weight-value-supply control unit;

processing analog output data output from the data output terminal for every analog multiply-accumulate operation performed by said neuron core as serial data and/or parallel data;

receiving and holding, at the at least one second register of the data-hold-supply control unit, the analog output data output from the data output terminal; and

supplying the analog output data held by the at least one second register of the data-hold-supply control unit to the data input terminal of the neuron core having the data output terminal for use in the subsequent analog multiply-accumulate operation by the neuron core,

wherein the method further comprises controlling timing of holding data at the second register of the data-hold-supply control unit and timing of opening and closing at the path, and controlling the selecting operation of the second selector with respect to the input data from the prior stage and the hold data in the second registers.

7. The method of control of a neural network according to claim 6 , further comprising referring to data and a weight value set at said weight-value-supply control unit and data set at said data-hold-supply control unit.

8. The method of control of a neural network according to claim 6 , further comprising comparing data set at said data-hold-supply control unit with analog output data from the data output terminal and controlling the state of the neural network based on the comparison results.

9. A processor of a neural network,

the neural network comprising:

at least one neuron core having a data input terminal for receiving analog input data, a data output terminal for outputting analog output data, and a weight value input terminal, the neuron core being configured to perform an analog multiply-accumulate operation based on the analog input data arriving from the data input terminal and a weight value arriving from the weight value input terminal, the weight value being analog data;

a weight-value-supply control unit connected to the weight value input terminal and supplying the weight value from the weight value input terminal to said neuron core; and

a data-hold-supply control unit connected to the data input terminal and the data output terminal, and provided with at least one second register and a second selector, the second register being configured to hold at least the analog output data operated by the neuron core and output from the data output terminal such that the hold data is suppliable to output to the second selector, and the second selector being capable of selecting any one of the at least one second register or an input data arriving from a prior stage, and wherein the second selector causes opening and closing a path for transmitting the held data in the second register to the data input termina,

wherein, the processor is configured to:

control the supply of the weight value from said weight-value-supply control unit;

process the analog output data output from the data output terminal at every analog multiply-accumulate operation performed by said neuron core as serial data and/or parallel data;

receive and hold, at the at least one second register of the data-hold-supply control unit, the analog output data output from the data output terminal; and

supply the analog output data held by the at least one second register of the data-hold-supply control unit to the data input terminal of the neuron core having the data output terminal for use in the subsequent analog multiply-accumulate operation by the neuron core,

wherein the processor is further configured to control timing of holding data at the second register of said data-hold-supply control unit and timing of opening and closing at the path, and controlling the selecting operation of the second selector with respect to the input data from the prior stage and the hold data in the second registers.

10. The processor of a neural network according to claim 9 , wherein said weight-value-supply control unit comprises

a plurality of first registers respectively holding weight values and

a first selector for selecting any one of the plurality of registers and opening and closing a path for transmitting the held weight value to the weight value input terminal.