JP7241563B2 - neural network circuit - Google Patents

Description

The present invention relates to neural network circuits.

A neural network is generally configured using a sum-of-products operation circuit. In addition, since the neural network has a different configuration depending on the object to be learned or inferred (hereinafter referred to as the object to be learned or inferred), when changing the object to be learned or inferred, the neural network configuration changes, It is necessary to change the configuration of the neural network circuit. Therefore, conventionally, when changing the learning or inference target of the neural network, the reconfiguration function of FPGA (field-programmable gate array) is used to rewrite the logic circuit in FPGA (for example, , see Patent Document 1). Note that the reconfiguration function of FPGA is to write changed design data into FPGA. Means the ability to rewrite.

JP 2018-132830 A

In the conventional method of rewriting the FPGA logic circuit using the reconfiguration function, it takes time to operate the reconfiguration function of the FPGA. There is a problem that it takes time to rebuild the configuration of the neural network circuit corresponding to the learning or inference target.
SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and aims to provide a technique capable of changing the configuration of a neural network circuit in a shorter time than the reconfiguration function of FPGA. aim.

A neural network circuit according to the present invention includes an arithmetic circuit that performs arithmetic processing on input data and outputs output data that is the result of the arithmetic processing, and an input data control unit that controls the supply of input data to the arithmetic circuit. and a plurality of multiplication units each calculating a product of the input data supplied by the input data control unit and the weight parameter stored in the storage unit. an adder that calculates the sum of the products calculated by the plurality of multipliers; and a selection unit that feeds back the sum calculated by the adder to the adder according to a selection parameter stored in the storage unit. , the input data control unit supplies input data to some or all of the multiple multipliers, and some or all of the multiple multipliers each have an input data control unit calculates the product of the input data supplied by and the weight parameter stored in the storage unit, and the adder calculates the products calculated by some or all of the multiple multipliers using the input data, respectively. The input data control unit supplies input data different from the input data to some or all of the multiple multipliers, and A part or all of the multipliers of each of the multipliers calculates a product of another input data supplied by the input data control unit and a weight parameter stored in the storage unit, and an adder of the plurality of multipliers Some or all of the multipliers calculate a second sum that is a sum of products calculated using different input data, and the selection unit selects the first sum calculated by the adders according to the selection parameter is fed back to the adder, and the adder calculates the third sum that is the sum of the first sum fed back by the selector and the calculated second sum.
A neural network circuit according to the present invention includes an arithmetic circuit that performs arithmetic processing on input data and outputs output data that is the result of the arithmetic processing, and an input data control unit that controls the supply of input data to the arithmetic circuit. and a plurality of multiplication units each calculating a product of the input data supplied by the input data control unit and the weight parameter stored in the storage unit. an adder that calculates the sum of the products calculated by the plurality of multipliers; and a selection unit that feeds back the sum calculated by the adder to the adder according to a selection parameter stored in the storage unit. , the input data control unit supplies first input data to a first multiplier of the plurality of multipliers and supplies second input data to a second multiplier of the plurality of multipliers; a first multiplier for calculating a first product of the first input data supplied by the input data control unit and a weight parameter stored by the storage unit; and a second multiplier for , the second input data supplied by the input data control unit and the weight parameter stored in the storage unit to calculate a second product, the adder calculating the product of the first product and the second product; The input data control unit supplies the third input data to a first multiplier of the plurality of multipliers, and the fourth input data to the plurality of multiplications. The first multiplier supplies the third product which is the product of the third input data provided by the input data control unit and the weight parameter stored in the storage unit. , the second multiplier calculates a fourth product of the fourth input data supplied by the input data control unit and the weight parameter stored in the storage unit, and the adder calculates the fourth Calculate a second sum that is the sum of the product of 3 and the fourth product, the selection unit feeds back the first sum calculated by the adder to the adder according to the selection parameter, and the adder , a third sum that is the sum of the first sum fed back by the selector and the second sum calculated.
A neural network circuit according to the present invention includes an arithmetic circuit that performs arithmetic processing on input data and outputs output data that is the result of the arithmetic processing, and an input data control unit that controls the supply of input data to the arithmetic circuit. and a plurality of multiplication units each calculating a product of the input data supplied by the input data control unit and the weight parameter stored in the storage unit. an adder that calculates the sum of the products calculated by the plurality of multipliers; and a selection unit that feeds back the sum calculated by the adder to the adder according to a selection parameter stored in the storage unit. , a plurality of arithmetic circuits, wherein a selection unit included in a first arithmetic circuit among the plurality of arithmetic circuits selects a second arithmetic circuit among the plurality of arithmetic circuits in accordance with a selection parameter stored in the storage unit; further outputs the sum calculated by the adder provided in to the adder provided in the first arithmetic circuit, and the adder provided in the first arithmetic circuit outputs the calculated sum and the second calculation Calculate the sum with the sum calculated by the adder provided in the circuit.

The configuration of the neural network circuit can be changed in a shorter time than the reconfiguration function of FPGA.

1 is a block diagram showing the configuration of a learning/inference system including a neural network circuit according to Embodiment 1; FIG. 1 is a block diagram showing the configuration of a neural network circuit according to Embodiment 1; FIG. 4 is a flow chart showing a first arithmetic processing method by the neural network circuit according to Embodiment 1; 4 is a diagram for explaining a first specific example of the operation of the neural network circuit according to Embodiment 1; FIG. 7 is a flow chart showing a second arithmetic processing method by the neural network circuit according to Embodiment 1; FIG. 9 is a diagram for explaining a second specific example of the operation of the neural network circuit according to Embodiment 1; 8 is a block diagram showing the configuration of a neural network circuit according to Embodiment 2; FIG. 9 is a flowchart showing an arithmetic processing method by a neural network circuit according to Embodiment 2; FIG. 10 is a diagram for explaining a specific example of the operation of the neural network circuit according to the second embodiment;

Hereinafter, in order to describe the present invention in more detail, embodiments for carrying out the present invention will be described with reference to the accompanying drawings.
Embodiment 1.
FIG. 1 is a block diagram showing the configuration of a learning/inference system 1 including a neural network circuit 20 according to Embodiment 1. As shown in FIG. As shown in FIG. 1, the learning/inference system 1 includes an information input section 10, a neural network circuit 20, a control section 30, and a learning/inference result output section .

The information input unit 10 acquires input data for learning or inference by the neural network circuit 20 through user input. The information input section 10 outputs input data to the neural network circuit 20 via the control section 30 . Examples of the information input unit 10 include a camera, microphone, keyboard, and the like. Examples of input data input by the information input unit 10 include image data, voice data, character string data, and the like.
The neural network circuit 20 refers to the input data acquired from the information input unit 10 and performs learning processing and inference processing. The neural network circuit 20 outputs output data resulting from the learning process and the inference process to the control unit 30 . A detailed description of the neural network circuit 20 will be given later.

The control unit 30 refers to the input data acquired from the information input unit 10, and sets or updates weighting parameters and selection parameters used in learning processing and inference processing. Next, the control unit 30 outputs the input data, the weight parameter and the selection parameter to the neural network circuit 20 and acquires output data from the neural network circuit 20 . Next, the control unit 30 refers to the output data to set or update the weight parameters and selection parameters used by the neural network circuit 20 for learning processing and inference processing. Through the steps described above, the control unit 30 controls the learning processing and the inference processing by the neural network circuit 20 . Note that the control unit 30 may reference the output data acquired from the neural network circuit 20 and return the output data to the neural network circuit 20 after updating the weight parameter. In that case, the neural network circuit 20 may perform learning processing and inference processing again using the output data. In addition, the control unit 30 converts the output data acquired from the neural network circuit 20 into a form suitable for output by the learning/inference result output unit 40, and outputs the converted data to the learning/inference result output unit 40. . Examples of the converted data include image data, audio data, character data, and the like. An example of the control unit 30 is a CPU.
The learning/inference result output unit 40 outputs the converted data acquired from the control unit 30 in the form of images, voices, characters, or the like. Examples of the learning/inference result output unit 40 include a display, an LED, a speaker, and the like.

FIG. 2 is a block diagram showing the configuration of neural network circuit 20 according to the first embodiment. As shown in FIG. 2, the neural network circuit 20 includes an input data control section 200, a buffer 210, a first arithmetic circuit 800, and a second arithmetic circuit 900 having the same configuration as the first arithmetic circuit 800. ing. In Embodiment 1, a configuration in which neural network circuit 20 includes two arithmetic circuits will be described. can be a number.

The buffer 210 acquires input data from the control unit 30 and temporarily accumulates the input data. The buffer 210 outputs part or all of the accumulated input data to the input data control section 200 .
The input data control section 200 acquires a plurality of pieces of input data from the buffer 210 and controls supply of each input data to the first arithmetic circuit 800 and the second arithmetic circuit 900 . The details of how the input data control unit 200 controls the supply of input data will be described later. Further, the input data control unit 200 further controls each arithmetic processing by the first arithmetic circuit 800 and the second arithmetic circuit 900 by controlling the output of weight parameters and selection parameters from the storage unit 700, which will be described later. .

The first arithmetic circuit 800 includes a multiplier 300 , a multiplier 301 , a multiplier 302 , a selection section 400 , an adder 500 , an output section 600 and a storage section 700 . In this embodiment, a configuration in which each of the first arithmetic circuit 800 and the second arithmetic circuit 900 includes three multipliers will be described. However, the first arithmetic circuit 800 and the second arithmetic circuit 900 only need to have a plurality of multipliers. The number of multipliers included in circuit 900 may vary and is not limited to this configuration.

The storage unit 700 stores the weight parameters and selection parameters output by the control unit 30 . The storage unit 700 also outputs the stored weight parameter to the multiplier 300 , the multiplier 301 or the multiplier 302 based on the command from the input data control unit 200 . A weight parameter exists for each output multiplier, and these weight parameters may have the same value or different values. The storage unit 700 also outputs the stored selection parameters to the selection unit 400 based on the command from the input data control unit 200 . In this embodiment, the input data control unit 200 controls the output of the weighting parameters and the selection parameters from the storage unit 700. However, the control unit 30 outputs the weighting parameters and the selection parameters from the storage unit 700. may be controlled.

Multipliers 300 , 301 , and 302 each calculate the product of the input data supplied by the input data control unit 200 and the weight parameter output by the storage unit 700 . Multiplier 300 , multiplier 301 , and multiplier 302 each output the calculated product to adder 500 .
Adder 500 calculates the sum of the products calculated by multipliers 300 , 301 , and 302 . The adder 500 outputs the calculated sum to the selection section 400 or the output section 600 according to the operation of the selection section 400, which will be described later.

The selection unit 400 feeds back the sum calculated by the adder 500 to the adder 500 according to the selection parameter output by the storage unit 700 .
For example, if the selection parameter is a value that instructs the adder 500 to feed back the first sum calculated by the adder 500 , the selection unit 400 sends the first sum to the adder 500 . give feedback. Next, the adder 500 adds the first sum fed back by the selection unit 400 and the second sum, which is the sum calculated after calculating the first sum, to obtain the third sum, which is the addition result. is output to the output unit 600. On the other hand, if the selection parameter is a value that instructs not to feed back the first sum to the adder 500 , the selection unit 400 outputs a 0 value to the adder 500 . In that case, the adder 500 outputs the calculated sum to the output section 600 as it is. In other words, the output destination of the adder 500 is the output unit 600 when the calculated sum is fed back from the selection unit 400 and when a 0 value is obtained from the selection unit 400, and otherwise the selection unit 400 is.
The output unit 600 outputs the sum acquired from the adder 500 to the control unit 30 as output data.

The second arithmetic circuit 900 includes a multiplier 310 , a multiplier 311 , a multiplier 312 , a selection section 410 , an adder 510 , an output section 610 and a storage section 710 . The multiplier 310, the multiplier 311, the multiplier 312, the selector 410, the adder 510, the output unit 610, and the storage unit 710 are respectively the multiplier 310, the multiplier 311, the multiplier 312, the selector 400, and the storage unit 710 described above. It has the same functions as adder 500 , output section 600 and storage section 700 .

More specifically, the multiplier 310 calculates the product of the same input data as the input data supplied to the multiplier 300 by the input data control unit 200 and a weight parameter different from the weight parameter used by the multiplier 300. do. Multiplier 311 calculates the product of the same input data as the input data supplied to multiplier 301 by input data control section 200 and a weight parameter different from the weight parameter used by multiplier 301 . The multiplier 312 calculates the product of the same input data as the input data supplied to the multiplier 302 by the input data control unit 200 and a weight parameter different from the weight parameter used by the multiplier 302 .

The adder 510 calculates the sum of the products calculated by the multiplier 310, the multiplier 311, and the multiplier 312, and the selector 410 calculates the The resulting sum is fed back to adder 510 . The output unit 610 outputs the sum acquired from the adder 510 to the control unit 30 as output data. With the above configuration, the neural network circuit 20 outputs two types of output data, the output data calculated by the first arithmetic circuit 800 and the output data calculated by the second arithmetic circuit 900 . As described above, the number of arithmetic circuits included in the neural network circuit 20 is not limited to two, and may be any number. Therefore, the neural network circuit 20 according to this embodiment can output the same number of output data as the number of arithmetic circuits provided.

Next, operation of the neural network circuit 20 according to Embodiment 1 will be described with reference to the drawings. It should be noted that the neural network circuit 20 starts the operation when the input data, the weight parameter and the selection parameter are obtained from the control unit 30 . Further, since the operation of the first arithmetic circuit 800 and the operation of the second arithmetic circuit 900 are the same, the explanation of the operation of the arithmetic circuits included in the neural network circuit 20 will be given only for the first arithmetic circuit 800. Only the operation of is explained.

First, when the number of input data used by the neural network circuit 20 to calculate one output data is equal to or less than the number of multipliers provided in the first arithmetic circuit 800 and the second arithmetic circuit 900 , the operation of the neural network circuit 20 will be described. FIG. 3 is a flowchart showing an arithmetic processing method by the neural network circuit 20 according to the first embodiment in this case. Moreover, in the specification of the present application, the term “simultaneously” does not have to be exactly the same time, but includes substantially the same time.

As shown in FIG. 3, the input data control section 200 supplies input data to the multipliers 300, 301, and 302 (step ST1). The input data control unit 200 also instructs the storage unit 700 to output weight parameters to the multipliers 300, 301, and 302, respectively. The input data control unit 200 also commands the storage unit 700 to output the selection parameter to the selection unit 400 .

The multipliers 300, 301, and 302 each calculate the product of the input data supplied by the input data control section 200 and the weight parameter obtained from the storage section 700 (step ST2). Multiplier 300 , multiplier 301 , and multiplier 302 output each calculated product to adder 500 .

The adder 500 calculates the sum of the products calculated by the multipliers 300, 301, and 302 (step ST3).
The selection unit 400 outputs a 0 value to the adder 500 based on the selection parameter acquired from the storage unit 700 (step ST4). Next, the adder 500 outputs the sum calculated in step ST3 to the output section 600 as it is.
The output unit 600 outputs the sum acquired from the adder 500 to the control unit 30 as output data (step ST5).

Next, when the number of input data used by the neural network circuit 20 to calculate one output data is less than or equal to the number of multipliers provided in the first arithmetic circuit 800 and the second arithmetic circuit 900 A specific example of the operation of the neural network circuit 20 will be described. FIG. 4 is a diagram for explaining the specific example. In FIG. 4, each block described in the same row represents each data processed by any configuration provided in the neural network circuit 20, and each data described in the same column represents the neural network Each of the data is shown to be processed simultaneously by any of the components of circuit 20. FIG. In addition, shaded blocks in FIG. 4 indicate that the configuration of the neural network circuit 20 corresponding to the block does not perform processing, or that the configuration of the neural network circuit 20 corresponding to the block performs processing. Indicates that the processing result will not be output to another configuration even if it is executed. Further, in the following specific example, the number of input data used by the neural network circuit 20 to calculate one output data is 3, and the first arithmetic circuit 800 and the second arithmetic circuit 900 each include An example in which the number of multipliers is three will be described. is equal to or less than the number of multipliers provided, and is not limited to this example.

First, as shown in column A of FIG. 4, in step ST1 described above, input data control section 200 simultaneously supplies input data D11 to multiplier 300, supplies input data D12 to multiplier 301, The multiplier 302 is supplied with the input data D13. Further, input data control section 200 controls these parameters so that storage section 700 outputs weight parameter M11 to multiplier 300, weight parameter M12 to multiplier 301, and weight parameter M13 to multiplier 302. The storage unit 700 is instructed at the same time as the input data is supplied. After these steps, the multiplier 300 calculates the product of the input data D11 and the weight parameter M11 in step ST2 described above, and the multiplier 301 calculates the product of the input data D12 and the weight parameter M12. Then, the multiplier 302 calculates the product of the input data D13 and the weight parameter M13.

Next, as shown in the column of B in FIG. 4, the adder 500 adds D11×M11, which is the product calculated by the multiplier 300, and D12×, which is the product calculated by the multiplier 301, in step ST3 described above. The sum of M12 and D13×M13, which is the product calculated by the multiplier 302, is calculated. After the calculation of the sum by the adder 500 or in parallel with the calculation of the sum, the selector 400 outputs a 0 value to the adder 500 based on the selection parameter in step ST4 described above. Next, the adder 500 outputs the calculated sum D11×M11+D12×M12+D13×M13 as it is to the output section 600, and the output section 600 sends the sum as output data to the control section 30 in step ST5 described above. Output.
In parallel with the execution of the above steps by the adder 500, the input data control section 200 simultaneously supplies the input data D21 to the multiplier 300, supplies the input data D22 to the multiplier 301, and and supplies the input data D23 to the multiplier 302 . At the same time when these input data are supplied, the input data control unit 200 outputs the weight parameter M21 to the multiplier 300, the weight parameter M22 to the multiplier 301, the weight parameter M22 to the multiplier 302, and the weight parameter M22 to the multiplier 302. The storage unit 700 is instructed to output M23. Next, in step ST2 described above, the multiplier 300 calculates the product of the input data D21 and the weight parameter M21, the multiplier 301 calculates the product of the input data D22 and the weight parameter M22, and the multiplier 302 calculates the product of the input data D22 and the weight parameter M22. calculates the product of the input data D23 and the weight parameter M23.

Next, as shown in column C of FIG. 4, adder 500 adds D21×M21, which is the product calculated by multiplier 300, and D22×, which is the product calculated by multiplier 301, in step ST3 described above. The sum of M22 and D23×M23, which is the product calculated by the multiplier 302, is calculated. After the calculation of the sum by the adder 500 or in parallel with the calculation of the sum, the selection unit 400 sets the adder 500 to 0 based on the selection parameter acquired from the storage unit 700 in step ST4 described above. to output Next, the adder 500 outputs the calculated sum D21×M21+D22×M22+D23×M23 as it is to the output section 600, and the output section 600 sends the sum as output data to the control section 30 in step ST5 described above. Output.
In parallel with execution of the above steps by adder 500, input data control section 200 simultaneously supplies input data D31 to multiplier 300, supplies input data D32 to multiplier 301, and supplies input data D31 to multiplier 301 in step ST1 described above. and supplies the input data D33 to the multiplier 302 . At the same time when these input data are supplied, the input data control unit 200 outputs the weight parameter M31 to the multiplier 300 , the weight parameter M32 to the multiplier 301 , the weight parameter M32 to the multiplier 302 , and the weight parameter M32 to the multiplier 302 . The storage unit 700 is instructed to output M33. Next, in step ST2 described above, the multiplier 300 calculates the product of the input data D31 and M31, the multiplier 301 calculates the product of the input data D32 and M32, the multiplier 302 calculates the input data Calculate the product of D33 and M33.

As described above, the number of input data used by the neural network circuit 20 to calculate one piece of output data is less than or equal to the number of multipliers provided in the first arithmetic circuit 800 and the second arithmetic circuit 900. In some cases, the adder 500 outputs the sum to the output section 600 without feeding back the sum calculated by the adder 500 to the adder 500 as in the above specific example.

Next, when the number of input data used by the neural network circuit 20 to calculate one output data is greater than the number of multipliers provided in each of the first arithmetic circuit 800 and the second arithmetic circuit 900 , the operation of the neural network circuit 20 will be described. FIG. 5 is a flowchart showing an arithmetic processing method by the neural network circuit 20 according to the first embodiment in this case.

The input data control unit 200 supplies the first input data to the multiplier 300, which is the first multiplier, and supplies the second input data to the multiplier 301, which is the second multiplier (step ST10). The input data control unit 200 also instructs the storage unit 700 to output weight parameters to the multipliers 300 and 301, respectively.

The multiplier 300 calculates a first product of the first input data obtained from the input data control unit 200 and the weight parameter obtained from the storage unit 700, and the multiplier 301 is the input data control unit. A second product of the second input data obtained from 200 and the weight parameter obtained from the storage unit 700 is calculated (step ST11). Multiplier 300 outputs the calculated first product to adder 500 , and multiplier 301 outputs the calculated second product to adder 500 .

The adder 500 calculates a first sum of the first product calculated by the multiplier 300 and the second product calculated by the multiplier 301 (step ST12). Adder 500 outputs the calculated first sum to selection section 400 .

The input data control section 200 supplies the third input data to the multiplier 300 and supplies the fourth input data to the multiplier 301 (step ST13). The input data control unit 200 also instructs the storage unit 700 to output weight parameters to the multipliers 300 and 301, respectively. Note that the weight parameter is a weight parameter different from the weight parameter in steps ST10 and ST11, and may have different values or the same value. The input data control unit 200 also commands the storage unit 700 to output the selection parameter to the selection unit 400 .

Multiplier 300 calculates a third product of the third input data obtained from input data control section 200 and the weight parameter obtained from storage section 700, and multiplier 301 calculates the third product of the input data control section. A fourth product of the fourth input data obtained from 200 and the weight parameter obtained from storage unit 700 is calculated (step ST14). Multiplier 300 outputs the calculated third product to adder 500 , and multiplier 301 outputs the calculated fourth product to adder 500 .

The adder 500 calculates a second sum of the third product calculated by the multiplier 300 and the fourth product calculated by the multiplier 301 (step ST15).
The selection unit 400 feeds back the first sum obtained from the adder 500 to the adder 500 based on the selection parameter obtained from the storage unit 700 (step ST16).
The adder 500 calculates a third sum that is the sum of the first sum fed back by the selection section 400 and the calculated second sum (step ST17). Adder 500 outputs the calculated third sum to output section 600 .
The output section 600 outputs the third sum acquired from the adder 500 to the control section 30 as output data (step ST18).

Next, when the number of input data used by the neural network circuit 20 to calculate one output data is greater than the number of multipliers provided in each of the first arithmetic circuit 800 and the second arithmetic circuit 900 A specific example of the operation of the neural network circuit 20 will be described. FIG. 6 is a diagram for explaining the specific example. In FIG. 6, each block described in the same row represents each data processed by any configuration provided in the neural network circuit 20, and each data described in the same column represents the neural network Each of the data is shown to be processed simultaneously by any of the components of circuit 20. FIG. The shaded blocks in FIG. 6 indicate that the configuration of the neural network circuit 20 corresponding to the block does not perform processing, or that the configuration of the neural network circuit 20 corresponding to the block performs processing. Indicates that the processing result will not be output to another configuration even if it is executed. Further, in the following specific example, the number of input data used by the neural network circuit 20 to calculate one output data is four, and the first arithmetic circuit 800 and the second arithmetic circuit 900 each include An example in which the number of multipliers is three will be described. is greater than the number of multipliers provided, and is not limited to this example.

As shown in column D in FIG. 6, input data control section 200 simultaneously supplies input data D11 (first input data) to multiplier 300 and input data D11 (first input data) to multiplier 301 in step ST10 described above. D12 (second input data) is supplied. At the same time as supplying these input data, the input data control unit 200 causes the storage unit 700 to output the weight parameter M11 to the multiplier 300 and the weight parameter M12 to the multiplier 301. Command. Then, after each of the above steps, multiplier 300 calculates the product (first product) of input data D11 and weight parameter M11 in step ST11 described above, and multiplier 301 calculates input data D12 and weight A product (second product) with the parameter M12 is calculated.

As shown in column E in FIG. 6, the adder 500 converts D11×M11, which is the product calculated by the multiplier 300, and D12×M12, which is the product calculated by the multiplier 301, in step ST12 described above. A sum (first sum) is calculated. After the calculation of the sum by the adder 500 or in parallel with the calculation of the sum, the input data control section 200 supplies the input data D13 (third input data) to the multiplier 300 in step ST13 described above. , and supplies input data D14 (fourth input data) to the multiplier 301 . At the same time as supplying these input data, the input data control unit 200 instructs the storage unit 700 to output the weight parameter M13 to the multiplier 300 and the weight parameter M14 to the multiplier 301 . do. Next, in step ST14 described above, multiplier 300 calculates the product (third product) of input data D13 and weight parameter M13, and multiplier 301 calculates the product (third product) of input data D14 and weight parameter M14. 4th product).

As shown in column F in FIG. 6, the adder 500 converts D13×M13, which is the product calculated by the multiplier 300, and D14×M14, which is the product calculated by the multiplier 301, in step ST15 described above. A sum (second sum) is calculated. After the calculation of the sum by the adder 500, or in parallel with the calculation of the sum, the selection unit 400 calculates the D11×M11+D12×M12, which is the sum of the obtained values, is output to the adder 500. FIG. Then, in step ST17 described above, the adder 500 adds the sum of the first sum D11×M11+D12×M12 obtained from the selection unit 400 and the calculated second sum D13×M13+D14×M14 (third sum) Calculate Next, the adder 500 outputs the calculated third sum D11×M11+D12×M12+D13×M13+D14×M14 as it is to the output unit 600, and the output unit 600 outputs the third sum in step ST18 described above. is output to the control unit 30 as output data. On the other hand, input data control section 200 returns to the process of step ST10 described above, supplies input data D21 (first input data) to multiplier 300, and supplies input data D22 (second input data) to multiplier 301. ). At the same time as supplying these input data, the input data control unit 200 stores the Command. Note that the configuration of the G and H columns in FIG. 6 is the same as the configuration of the E and F columns, and thus the description thereof is omitted.

As described above, in the above specific example, the first arithmetic circuit 800 and the second arithmetic circuit 900 each have the number of input data used by the neural network circuit 20 to calculate one piece of output data. More than the number of multipliers. In this case, the selection unit 400 feeds back the sum calculated by the adder 500 to the adder 500, so that the adder 500 adds the two sums calculated at different times. In the above specific example, the input data control section 200 supplies input data to the first multiplier 300 and the second multiplier 301 of the plurality of multipliers provided in the first arithmetic circuit in step ST10. , the first multiplier 300 and the second multiplier 301 respectively calculate the first product and the second product in step ST11, but the configuration is not limited to this. The input data control unit 200 can supply data to some or all of the multiple multipliers included in the first arithmetic circuit. Further, some or all of the multiple multipliers can each calculate the product of the input data supplied by the input data control section 200 and the weight parameter stored in the storage section 700 .

In the above description of the operation of the neural network circuit 20 according to the first embodiment, the number of input data used by the neural network circuit 20 to calculate one output data is the same as that of the first arithmetic circuit 800 and that of the second arithmetic circuit 800. The operation of neural network circuit 20 and the number of input data used by neural network circuit 20 to calculate one piece of output data when the number of multipliers included in arithmetic circuit 900 is less than or equal to the first The operation of the neural network circuit 20 when the number of multipliers provided in the arithmetic circuit 800 and the second arithmetic circuit 900 is greater than that of the arithmetic circuit 800 and the second arithmetic circuit 900 has been described. In this way, the operation of the neural network circuit 20 is based on the number of input data used by the neural network circuit 20 to calculate one piece of output data, and the number of input data used by the first arithmetic circuit 800 and the second arithmetic circuit 900. It depends on the number of multipliers you have. Generalizing this, the input data control unit 200 sets the number of input data simultaneously supplied to each multiplier by the input data control unit 200 to m, and the first arithmetic circuit 800 and the second arithmetic circuit 900 are provided with m. Assuming that the number of multipliers connected is n, m data are temporarily stored in the buffer 210, ceil(m/ceil(m/n)) input data are acquired from the buffer 210, and n ceil(m/ceil(m/n)) pieces of input data are supplied to ceil(m/ceil(m/n)) pieces of multipliers among the multipliers. Note that "ceil()" in "ceil(m/ceil(m/n))" indicates a function that rounds up the argument to the first decimal place.

As described above, the neural network circuit 20 according to the first embodiment performs arithmetic processing on input data, and the first arithmetic circuit 800 that outputs the output data that is the result of the arithmetic processing; The first arithmetic circuit 800 includes a storage unit 700 for storing weighting parameters and selection parameters, and a storage unit 700 for storing the input data provided by the input data control unit 200. A multiplier 300, a multiplier 301, and a multiplier 302 of a plurality of multipliers that calculate the product of data and a weight parameter stored in the storage unit 700, and an addition that calculates the sum of the products calculated by the plurality of multipliers. and a selection unit 400 that feeds back the sum calculated by the adder 500 to the adder 500 according to the selection parameter stored in the storage unit 700 .

According to the above configuration, the number of input data used by the neural network circuit 20 to calculate one piece of output data is the number of multipliers provided in the first arithmetic circuit 800 and the second arithmetic circuit 900, respectively. If: . Further, according to the above configuration, the number of input data used by the neural network circuit 20 to calculate one output data is the number of multipliers provided in the first arithmetic circuit and the second arithmetic circuit. , the input data control unit 200 divides the input data and supplies the divided input data to a plurality of multipliers at different timings to perform time division of the input data, and the plurality of multipliers and , and the adder 500 sequentially process the divided input data, and the selector 400 feeds back the input data processed by the adder 500 in the past to the adder 500 . This enables time-division processing of input data. By controlling the supply of input data to a plurality of multipliers by the input data control unit 200 and by the feedback to the adder 500 by the selection unit 400, when changing the learning or inference target of the neural network, the FPGA The configuration of the neural network circuit can be changed without rewriting the logic circuit like the reconfiguration function of . Therefore, the configuration of the neural network circuit can be changed in a shorter time than the reconfiguration function of FPGA.

Moreover, according to the above configuration, as described above, the input data can be processed in parallel, so the time required for arithmetic processing can be reduced. In addition, according to the above configuration, as described above, input data can be time-divisionally processed. Therefore, the number of multipliers that perform parallel processing is limited, and arithmetic processing is performed even when the circuit scale is reduced. be able to. As a result, it is possible to dynamically change the balance between the arithmetic processing time of the neural network circuit and the circuit scale. Therefore, even if the object of learning or inference is changed, the optimal neural network circuit for the changed object of learning or inference can be constructed without taking much time.

In the neural network circuit 20 according to Embodiment 1, the input data control unit 200 supplies input data to some or all of the multiple multipliers, and supplies input data to one of the multiple multipliers. section or all the multipliers respectively calculate the product of the input data supplied by the input data control section 200 and the weight parameter stored in the storage section 700, and the adder 500 is one of the plurality of multipliers. or all of the multipliers calculate a first sum that is a sum of products calculated using the input data, and the input data control unit 200 controls some or all of the multiple multipliers to Input data different from the input data is supplied, and some or all of the multiple multipliers store the different input data supplied by the input data control unit 200 and the storage unit 700 A product with the weight parameter is calculated, and the adder 500 calculates a second sum that is a sum of products calculated by some or all of the multiple multipliers using different input data. Then, the selection unit 400 feeds back the first sum calculated by the adder 500 to the adder 500 according to the selection parameter, and the adder 500 feeds back the first sum fed back by the selection unit 400 and the calculated Calculate a third sum that is the sum with the second sum.

According to the above configuration, in the neural network circuit according to Embodiment 1, the input data control unit 200 divides the input data, and the divided input data is multiplied by some or all of the multiple multipliers. A part or all of the multiple multipliers and the adder 500 sequentially process the divided input data, and the selector 400 processes the input data in the past. The obtained data is fed back to the adder 500 concerned. The adder 500 adds the current processed data and the data fed back by the selector 400 . This enables time-division processing of input data.

Further, in the neural network circuit 20 according to Embodiment 1, the input data control unit 200 supplies the first input data to the first multiplier 300 among the multiple multipliers, and the second input data is supplied to the second multiplier 301 of the plurality of multipliers, and the first multiplier 300 receives the first input data supplied by the input data control unit 200 and the weight parameter stored in the storage unit 700. , and the second multiplier 301 calculates a first product that is the product of the second input data supplied by the input data control unit 200 and the weight parameter stored in the storage unit 700 . 2, the adder 500 calculates a first sum that is the sum of the first product and the second product, and the input data control unit 200 converts the third input data into a plurality of a first multiplier 300 of the multipliers and fourth input data to a second multiplier 301 of the plurality of multipliers, the first multiplier 300 having an input data control A third product, which is the product of the third input data supplied by the unit 200 and the weight parameter stored in the storage unit 700, is calculated, and the second multiplier 301 calculates the third input data supplied by the input data control unit 200. 4 and the weight parameter stored in the storage unit 700, and the adder 500 calculates a second sum that is the sum of the third product and the fourth product. and the selection unit 400 feeds back the first sum calculated by the adder 500 to the adder 500 according to the selection parameter, and the adder 500 calculates the first sum fed back by the selection unit 400 and A third sum, which is the sum with the calculated second sum, is calculated.

According to the above configuration, in the neural network circuit according to Embodiment 1, the input data control unit 200 outputs a plurality of the first input data, the second input data, the third input data and the fourth input data. The input data is divided, the divided input data is supplied to the first multiplier 300 and the second multiplier 301, respectively, and the first multiplier 300 and the second multiplier 301, the adder 500 and the sequentially processes the divided input data, and the selector 400 feeds back the data processed by the adder 500 in the past to the adder 500 . The adder 500 adds the current processed data and the data fed back by the selector 400 . This enables time-division processing of input data.

Further, in the neural network circuit 20 according to the first embodiment, the input data control unit 200 sets the number of input data to m, the number of multiple multipliers to n, and ceil(m/ ceil(m/n)) multipliers are supplied with ceil(m/ceil(m/n)) pieces of input data.
According to the above configuration, the number of input data used by the neural network circuit 20 to calculate one piece of output data is the number of multipliers provided in the first arithmetic circuit 800 and the second arithmetic circuit 900, respectively. Input data can be processed in parallel, as described above, if: Also, when the number of input data used by the neural network circuit 20 to calculate one output data is greater than the number of multipliers provided in the first arithmetic circuit 800 and the second arithmetic circuit 900, , time-division processing of the input data can be performed.

Embodiment 2.
In Embodiment 1, the configuration in which the first arithmetic circuit 800 performs arithmetic processing on input data and outputs output data as a result of the arithmetic processing has been described. Although not described in the first embodiment, the second arithmetic circuit 900 also performs arithmetic processing on input data in the same manner as the first arithmetic circuit 800, and outputs output data as a result of the arithmetic processing. . That is, in Embodiment 1, the first arithmetic circuit 800 and the second arithmetic circuit 900 independently perform arithmetic processing on input data and output output data as a result of the arithmetic processing. On the other hand, in the second embodiment, a configuration is described in which the adder 501 of the first arithmetic circuit 800 can further use the sum calculated by the adder 511 of the second arithmetic circuit 901 to perform arithmetic processing. .

Embodiment 2 will be described below with reference to the drawings. In addition, the same reference numerals are given to the configurations having the same functions as the configurations described in the first embodiment, and the description thereof will be omitted.
FIG. 7 is a block diagram showing the configuration of neural network circuit 21 according to the second embodiment. As shown in FIG. 7, the neural network circuit 21 according to the second embodiment differs from the neural network circuit 20 according to the first embodiment in that the input data control unit 201 includes a multiplier 300, a multiplier 301, a multiplier 302, multiplier 310, multiplier 311 and multiplier 312 are individually connected. Thereby, the input data control section 201 can make the input data supplied to the multiplier 300 and the input data supplied to the multiplier 310 the same or different. Also, the input data control section 201 can make the input data supplied to the multiplier 301 and the input data supplied to the multiplier 311 different. Also, the input data control section 201 can make the input data supplied to the multiplier 302 and the input data supplied to the multiplier 312 different.

Further, the neural network circuit 21 according to the second embodiment differs from the neural network circuit 20 according to the first embodiment in that the output side of the adder 511 of the second arithmetic circuit 901 is connected to the selection of the first arithmetic circuit 801. The point is that it is connected to the unit 401 . Thereby, the adder 511 outputs the calculated sum to the selection unit 401 . The selection unit 401 outputs the sum calculated by the adder 511 to the adder 501 according to the selection parameter acquired from the storage unit 700 . The adder 501 calculates the sum of the calculated sum and the sum obtained from the selection unit 401 and outputs the sum to the output unit 600 .

Next, an arithmetic processing method by the neural network circuit 21 according to Embodiment 2 will be described with reference to the drawings. FIG. 8 is a flow chart showing an arithmetic processing method by the neural network circuit 21 according to the second embodiment. Steps ST20 to ST22 of the arithmetic processing method by the neural network circuit 21 according to the second embodiment are the same as steps ST1 to ST3 of the arithmetic processing method according to the first embodiment, respectively. Therefore, description of steps ST20 to ST22 is omitted.

Moreover, each process of step ST21 and step ST22 shall be performed in parallel by the 1st arithmetic circuit 801 and the 2nd arithmetic circuit 901. FIG. Further, in the arithmetic processing method described below, the input data control unit 201 supplies different input data to the multipliers 300 and 310, supplies different input data to the multipliers 301 and 311, This is an example in which different input data are supplied to the multipliers 302 and 312, respectively, and only the output section 600 outputs the output data.

First, after the process of step ST22, the adder 511 outputs the sum calculated in step ST22 to the selection section 401 of the first arithmetic circuit 801. FIG. Next, based on the selection parameter acquired from the storage unit 700, the selection unit 401 of the first arithmetic circuit 801 outputs the sum calculated in step ST22 by the adder 511 of the second arithmetic circuit 901 to the first arithmetic circuit. It is output to the adder 501 of 801 (step ST23).

Next, the adder 501 of the first arithmetic circuit 801 calculates the sum of the sum calculated in step ST22 and the sum calculated by the adder 511 (step ST24). The adder 501 outputs the calculated sum to the output unit 600 .
Next, the output unit 600 outputs the sum calculated by the adder 501 in step ST24 to the control unit 30 as output data (step ST25).

Either or both of the first arithmetic circuit 801 and the second arithmetic circuit 901 of the neural network circuit 21 according to the second embodiment are described in the first embodiment instead of steps ST20 to ST22. Steps ST10 to ST17 may be performed. When the second arithmetic circuit 901 performs steps ST10 to ST17 described in the first embodiment, the third sum calculated by the adder 511 in step ST17 is calculated by the second arithmetic circuit 901 in step ST23. The adder 511 is used instead of the sum calculated in step ST22.

Further, in the above description of the arithmetic processing method, the input data control unit 201 supplies different input data to the multipliers 300 and 310, supplies different input data to the multipliers 301 and 311, An example of supplying different input data to the multiplier 302 and the multiplier 312 and outputting the output data from the output unit 600 has been described. On the other hand, the input data control unit 201 supplies the same input data to the multipliers 300 and 310, supplies the same input data to the multipliers 301 and 311, and supplies the same input data to the multipliers 302 and 311. 312 are supplied with the same input data, and the output unit 600 and the output unit 610 respectively output the output data, a method similar to the arithmetic processing method according to the first embodiment can be performed. More specifically, instead of step ST23 described above, the selection unit 401 of the first arithmetic circuit 801 causes the adder 511 of the second arithmetic circuit 901 to perform A zero value is output to the adder 501 of the first arithmetic circuit 801 instead of the calculated sum. Next, instead of steps ST24 and ST25, the output unit 600 outputs the sum calculated by the adder 501 as output data to the control unit 30 as it is, and the output unit 610 outputs the sum calculated by the adder 511 in step ST22. The sum is directly output to the control unit 30 as output data.

Next, a specific example of operation of the neural network circuit 21 according to the second embodiment will be described. FIG. 9 is a diagram for explaining the specific example. In FIG. 9, each block described in the same row represents each data processed by any configuration provided in the neural network circuit 21, and each data described in the same column represents the neural network Each of the data is shown to be processed simultaneously by any component of circuit 21 . In addition, shaded blocks in FIG. 9 indicate that the configuration of the neural network circuit 20 corresponding to the block does not perform processing, or that the configuration of the neural network circuit 20 corresponding to the block performs processing. Indicates that the processing result will not be output to another configuration even if it is executed.

As shown in column I in FIG. 9, input data control section 201 simultaneously supplies input data D14 to multiplier 310, supplies input data D15 to multiplier 311, and supplies input data D15 to multiplier 311 in step ST20 described above. 312 with input data D16. At the same time when these input data are supplied, the input data control unit 201 outputs the weight parameter M14 to the multiplier 310, the weight parameter M15 to the multiplier 311, the weight parameter M15 to the multiplier 312, and the weight parameter M15 to the multiplier 312. The storage unit 700 is instructed to output M16. After these steps, multiplier 310 calculates the product of input data D14 and weight parameter M14 in step ST21 described above, and multiplier 301 calculates the product of input data D15 and weight parameter M15. Then, the multiplier 302 calculates the product of the input data D16 and the weight parameter M16.

Next, as shown in column J in FIG. 9, adder 511 adds D14×M14, which is the product calculated by multiplier 310, and D15×, which is the product calculated by multiplier 301, in step ST22 described above. The sum of M15 and D16×M16, which is the product calculated by the multiplier 302, is calculated. Next, the selection unit 401 of the first arithmetic circuit 801 selects the sum calculated in step ST22 by the adder 511 of the second arithmetic circuit 901 based on the selection parameter acquired from the storage unit 700 in step ST23 described above. A certain D14×M14+D15×M15+D16×M16 is output to the adder 501 of the first arithmetic circuit 801. FIG. On the other hand, in step ST20 described above, the input data control section 201 simultaneously supplies input data D11 to the multiplier 300, supplies input data D12 to the multiplier 301, and supplies input data D13 to the multiplier 302. . In step ST20 described above, the input data control unit 201 supplies the input data D24 to the multiplier 310, supplies the input data D25 to the multiplier 311, and supplies the input data D25 to the multiplier 312 at the same time as supplying these input data. Data D26 is supplied. In addition, input data control section 201 controls these parameters so that storage section 700 outputs weight parameter M11 to multiplier 300 , weight parameter M12 to multiplier 301 , and weight parameter M13 to multiplier 302 . At the same time as supplying the input data, the storage unit 700 is commanded. Further, input data control section 201 controls these parameters so that storage section 710 outputs weight parameter M24 to multiplier 310, weight parameter M25 to multiplier 311, and weight parameter M26 to multiplier 302. At the same time as supplying the input data, the storage unit 700 is commanded. Next, in step ST21 described above, multiplier 300 calculates the product of input data D11 and weight parameter M11, multiplier 301 calculates the product of input data D12 and weight parameter M12, and multiplier 302 calculates the product of input data D12 and weight parameter M12. calculates the product of the input data D13 and the weight parameter M13. In step ST21 described above, the multiplier 310 calculates the product of the input data D24 and the weight parameter M24, the multiplier 311 calculates the product of the input data D25 and the weight parameter M25, and the multiplier 312 calculates the product of the input data D25 and the weight parameter M25. , the product of the input data D26 and the weight parameter M26 is calculated.

Next, as shown in column K in FIG. 9, the adder 501 of the first arithmetic circuit 801 converts D11×M11, which is the product calculated by the multiplier 300 in step ST22 described above, and the multiplier 301 The sum of the calculated product D12×M12 and the product calculated by the multiplier 302 D13×M13 is calculated. Next, in step ST24 described above, adder 501 calculates the sum of this sum and D14×M14+D15×M15+D16×M16, which is the sum output from selection section 401 to adder 501 in step ST23. In step ST25 described above, the output unit 600 outputs D11×M11+D12×M12+D13×M13+D14×M14+D15×M15+D16×M16, which is the sum calculated in step ST24 by the adder 501, to the control unit 30 as output data. On the other hand, in step ST20 described above, the input data control section 201 simultaneously supplies input data D21 to the multiplier 300, supplies input data D22 to the multiplier 301, and supplies input data D23 to the multiplier 302. . In step ST20 described above, the input data control unit 201 supplies input data D34 to the multiplier 310, supplies input data D35 to the multiplier 311, and supplies input data D35 to the multiplier 312 at the same time as supplying these input data. Data D36 is supplied. Further, input data control section 201 controls these parameters so that storage section 700 outputs weight parameter M21 to multiplier 300, weight parameter M22 to multiplier 301, and weight parameter M23 to multiplier 302. At the same time as supplying the input data, the storage unit 700 is commanded. Further, input data control section 201 controls these parameters so that storage section 710 outputs weight parameter M34 to multiplier 310, weight parameter M35 to multiplier 311, and weight parameter M36 to multiplier 302. At the same time as supplying the input data, the storage unit 700 is commanded. In step ST22 described above, the adder 510 supplies the input data from the input data control unit 201 and the instruction to the storage unit 700 at the same time as D24×M24, which is the product calculated by the multiplier 310, and the multiplication The sum of D25×M25, which is the product calculated by the multiplier 311, and D26×M26, which is the product calculated by the multiplier 312, is calculated. Note that the configuration regarding the L column in FIG. 9 is the same as the configuration regarding the K column, and thus description thereof is omitted.

As described above, the neural network circuit 20 according to Embodiment 1 includes a plurality of arithmetic circuits, and the input data control unit 201 controls the supply of each input data to the plurality of arithmetic circuits. The selection unit 401 provided in the first arithmetic circuit 801 among the plurality of arithmetic circuits selects the adder 511 provided in the second arithmetic circuit 901 according to the selection parameter stored in the storage unit 700. The calculated sum is further output to the adder 501 provided in the first arithmetic circuit 801, and the adder 501 provided in the first arithmetic circuit 801 outputs the calculated sum and the second arithmetic circuit 901 The sum with the sum calculated by the adder 511 provided is calculated.

According to the above configuration, the first arithmetic circuit 801 and the second arithmetic circuit 901 process the input data in parallel, thereby operating the neural network circuit 21 so as to shorten the arithmetic processing time. can. Further, the selection unit 401 outputs the sum calculated by the adder 511 according to the selection parameter to the adder 501, and the adder 501 of the first arithmetic circuit 801 outputs the calculated sum and the sum of the second arithmetic circuit 901. By calculating the sum of the sums calculated by the adder 511, it is possible to increase the number of pieces of input data that are subjected to arithmetic processing in parallel in order to output one piece of output data. This can improve the performance of the learning or inference target.
In addition, within the scope of the present invention, it is possible to freely combine each embodiment, modify any component of each embodiment, or omit any component in each embodiment. .

1 learning/inference system 10 information input unit 20 neural network circuit 21 neural network circuit 30 control unit 40 learning/inference result output unit 200 input data control unit 201 input data control unit 210 buffer 300 multiplication unit, 301 multiplier, 302 multiplier, 310 multiplier, 311 multiplier, 312 multiplier, 400 selection unit, 401 selection unit, 410 selection unit, 500 adder, 501 adder, 510 adder, 511 adder, 600 output unit, 610 output unit, 700 storage unit, 710 storage unit, 800 first arithmetic circuit, 801 first arithmetic circuit, 900 second arithmetic circuit, 901 second arithmetic circuit.

Claims (4)

an arithmetic circuit that performs arithmetic processing on input data and outputs output data that is the result of the arithmetic processing;
an input data control unit that controls supply of input data to the arithmetic circuit;
The arithmetic circuit is
a storage unit for storing weight parameters and selection parameters;
a plurality of multipliers each calculating a product of input data supplied by the input data control unit and a weight parameter stored by the storage unit;
an adder that calculates the sum of the products calculated by the plurality of multipliers;
a selection unit that feeds back the sum calculated by the adder to the adder according to the selection parameter stored in the storage unit ;
The input data control unit supplies the input data to some or all of the multiple multipliers,
Some or all of the multiple multipliers calculate the product of the input data supplied by the input data control unit and the weight parameter stored in the storage unit, respectively;
the adder calculates a first sum that is a sum of products calculated by some or all of the multipliers using the input data;
The input data control unit supplies input data different from the input data to some or all of the multiple multipliers,
Some or all of the multiple multipliers each calculate a product of another input data supplied by the input data control unit and a weight parameter stored in the storage unit;
the adder calculates a second sum that is a sum of products calculated by some or all of the multipliers using the different input data;
The selection unit feeds back the first sum calculated by the adder to the adder according to the selection parameter,
The neural network circuit, wherein the adder calculates a third sum that is the sum of the first sum fed back by the selector and the calculated second sum. an arithmetic circuit that performs arithmetic processing on input data and outputs output data that is the result of the arithmetic processing;
an input data control unit that controls supply of input data to the arithmetic circuit;
The arithmetic circuit is
a storage unit for storing weight parameters and selection parameters;
a plurality of multipliers each calculating a product of input data supplied by the input data control unit and a weight parameter stored by the storage unit;
an adder that calculates the sum of the products calculated by the plurality of multipliers;
a selection unit that feeds back the sum calculated by the adder to the adder according to the selection parameter stored in the storage unit;
The input data control unit supplies first input data to a first multiplier of the plurality of multipliers, and supplies second input data to a second multiplier of the plurality of multipliers. supply to the vessel,
The first multiplier calculates a first product of the first input data supplied by the input data control unit and a weight parameter stored in the storage unit;
the second multiplier calculates a second product of the second input data supplied by the input data control unit and the weight parameter stored in the storage unit;
The adder calculates a first sum that is the sum of the first product and the second product;
The input data control unit supplies third input data to a first multiplier among the plurality of multipliers, and supplies fourth input data to a second multiplier among the plurality of multipliers. supply to the vessel,
the first multiplier calculates a third product of the third input data supplied by the input data control unit and the weight parameter stored in the storage unit;
the second multiplier calculates a fourth product of the fourth input data supplied by the input data control unit and a weight parameter stored in the storage unit;
The adder calculates a second sum that is the sum of the third product and the fourth product;
The selection unit feeds back the first sum calculated by the adder to the adder according to the selection parameter,
The neural network circuit, wherein the adder calculates a third sum that is the sum of the first sum fed back by the selector and the calculated second sum. an arithmetic circuit that performs arithmetic processing on input data and outputs output data that is the result of the arithmetic processing;
an input data control unit that controls supply of input data to the arithmetic circuit;
The arithmetic circuit is
a storage unit for storing weight parameters and selection parameters;
a plurality of multipliers each calculating a product of input data supplied by the input data control unit and a weight parameter stored by the storage unit;
an adder that calculates the sum of the products calculated by the plurality of multipliers;
a selection unit that feeds back the sum calculated by the adder to the adder according to the selection parameter stored in the storage unit;
comprising a plurality of the arithmetic circuits,
The selection unit provided in the first arithmetic circuit among the plurality of arithmetic circuits is provided in the second arithmetic circuit among the plurality of arithmetic circuits according to the selection parameter stored in the storage unit. further outputting the sum calculated by the adder to the adder included in the first arithmetic circuit;
The neural network circuit, wherein the adder provided in the first arithmetic circuit calculates the sum of the calculated sum and the sum calculated by the adder provided in the second arithmetic circuit. Where m is the number of input data and n is the number of the plurality of multipliers, the input data control unit controls ceil(m/ceil(m/n)) multipliers among the plurality of multipliers. 4. The neural network circuit according to any one of claims 1 to 3, wherein ceil(m/ceil(m/n)) pieces of said input data are supplied to each.

Images