JPH09319724A - Inference device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an inference device utilizing a neural network of a hierarchical structure type with the optional number of layers. SOLUTION: This device is provided with a single-layer neuron layer 12, a number of layers storing means, a synapse storing means, an input data storing means 13, an output data storing means 14, a data shifting means 24 shifting contents stored in the storing means 14 to the storing means 13, and a control means 23 weighting a weight stored in the synapse storing means to a synapse on the input/output side of the layer 12. Then, the device is constituted of an inference control means 26 repeating operation by the shifting means 24 and the control means 23 until the number of the layers becomes one stored in the number of layers storing means, and making the inference output of the neural network based on the output of the layer 12 at the time of amounting to the number of layers stored in the number of layers storing means.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a reasoning device using a neural network for reasoning using living nerve cells (hereinafter referred to as neurons) as a model.

[0002]

Hitherto, a hierarchical structure type is known as one of inference devices using a neural network. In this structure, usually three or more layers are formed in a hierarchical structure such as an input layer, one or more intermediate layers, and an output layer, and each layer is composed of a plurality of neurons. The layers are connected to each other by weighted synapses, and the weight of the synapse is determined by learning based on the learning data.

When the input data is input from the input layer when inferring, the information is transmitted to the neurons in the next layer through the synapse, and sequentially transmitted to the next layer by the firing phenomenon of the neuron, and finally. The output from the output layer is obtained as the inference result of the target model inferred in the neural network.

[0004]

Although there is such an inference apparatus in the related art, if the model to be inferred is large in the related art, the neural network itself is correspondingly enlarged, and together with that, There is a problem that the system becomes expensive. Even if the neural network becomes large (the number of layers increases), it has been desired to realize a small-sized and inexpensive neural network.

Further, in the conventional system, the configuration of the neural network is predetermined, and once constructed, the total number of neural networks cannot be increased or decreased. If the values do not match, inconvenience occurred. For example, if the size of the neural network is small with respect to the model to be inferred, inconvenience may occur, but if the neural network is large with respect to the model to be inferred, it becomes an uneconomically expensive inference device. There is. In particular, since the neural network is expensive, it has been desired to be able to flexibly deal with the model to be inferred.

Therefore, an object of the present invention is to provide a reasonably priced inference device for a neural network by forming a layer having neurons itself by only one layer. Another object of the present invention is to provide an inference device for a neural network that is composed of only one layer and that can arbitrarily change the number of layers.

[0007]

In order to achieve the above object, the invention according to claim 1 stores a single neuron layer which constitutes one layer of a plurality of neurons and the number of layers of a neural network which executes inference. Layer number storage means, synapse storage means for storing synapse weights corresponding to each layer of the neural network, input data storage means for storing data provided to the input side of the one neuron layer, and one neuron layer Output data storage means for storing data on the output side, data shift means for shifting the contents stored in the output data storage means to the input data storage means, and the synapse storage means corresponding to the number of layers being executed. And a weighting control means for weighting the synapses on the input and output sides of the one-layer neuron layer with the weights stored in , The operation by the data shift means and the weighting control means is repeated until the number of layers stored in the layer number storage means is reached, and the layer is stored when the number of layers stored in the layer number storage means is reached. An inference device, comprising inference control means for making an inference output of a neural network based on an output on the output side of a neuron layer.

According to the first aspect of the present invention, a one-layer neuron layer, an input data storage unit, and an output data storage unit are provided, and a synapse is provided between the one-layer neuron layer and the input data storage unit to execute synapse weighting. While setting and changing the weighting corresponding to the layers, the contents of the output data storage means are shifted to the input data storage means, and the operation of the hierarchical neural network composed of a large number of layers in one neuron layer is performed. .

The invention according to claim 2 corresponds to a single neuron layer forming a single layer of a plurality of neurons, a layer number storage means for storing the number of layers of a neural network for executing inference, and each layer of the neural network. Synapse storage means for storing synapse weights, input data storage means for storing data to be provided to the input side of the one-layer neuron layer, output data storage means for storing data on the output side of the one-layer neuron layer, and Data shift means for shifting the content stored in the output data storage means to the input data storage means, and the weights stored in the synapse storage means corresponding to the number of layers being executed are input / output to / from the one neuron layer Weighting control means for weighting the synapses on the side, and shift timing and shift timing by the data shift means. The inference synchronization means for timing the weighting of the synapses on the input / output side of the one-layer neuron layer by the weighting control means and the timing by the inference synchronization means until the number of layers stored in the layer number storage means is reached. Based on the output on the output side of the one-layer neuron layer when the number of layers stored in the layer number storage means is repeated, the inference output of the neural network is repeated based on the operations of the data shift means and the weight control means. And the inference control means.

According to the second aspect of the present invention, a single-layer neuron layer, an input data storage means, and an output data storage means are provided, a synapse is provided between the single-layer neuron layer and the input data storage means, and synapse weighting is executed. While setting and changing the weighting corresponding to the layers, the contents of the output data storage means are stored in the input data storage means by the synchronization timing based on the inference synchronization means until the maximum number of layers stored in the maximum layer number storage means is reached. As a result of shifting, a neural network of a hierarchical structure composed of many layers in one neuron layer is operated.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is an explanatory diagram showing the configuration of the inference apparatus of the present invention. As shown in FIG. 1, the inference apparatus 1 includes an execution unit 10, a control unit 20, a storage unit 30 and the like.

First, the execution unit 10 will be described. The execution unit 10 is composed of a single neuron layer 12, an input data storage means 13, and an output data storage means 14, and the input data storage means 13 and a single neuron layer. 1
A synapse 15 is connected between the two. The one-layer neuron layer 12 is only one layer of the hierarchical neural network.
2 has a plurality of neurons 11 arranged in a vertical position.

The input data storage means 13 stores a data output value to be input to the one-layer neuron layer 12 through the synapse 15, and outputs the data output value to the synapse 15 based on the stored data. The output data storage means 14 is for temporarily storing the output state of the one-layer neuron layer 12.

Next, the control unit 20 will be described. The control unit 20 includes an initial input control unit 21, an output control unit 22,
The weighting control means 23, the data shift means 24, the inference synchronization means 25, the inference control means 26, and the other control means 27 are included.

The initial input control means 21 is means for supplying data to the input layer (first layer) of the neural network for performing inference. The output control unit 22 is a unit for obtaining data from the output layer (final layer) of the neural network that performs inference. The weight control unit 23 is a unit for weighting the synapses 15 corresponding to the number of layers. The weight corresponding to the layer is stored in the weight storage unit 31.

The data shift means 24 shifts the contents of the output data storage means 14 to the input data storage means 13 in response to each increase in the number of execution layers when executing the inference in the execution section 10. It controls the movement of the transfer. The inference synchronization means 25 responds to the increase in the number of execution layers during the inference operation by the timing at which the weighting control means 23 weights the synapses 15, and the data shift means 24 by the storage contents of the output data storage means 14. It is a means for obtaining a synchronization timing such as a timing for shifting the contents of the input data storage means 13. Inference control means 26
Controls the overall operation related to inference, such as the mutual relationship between the above-mentioned control means.

Next, the storage unit 30 will be described. The storage unit 30 includes a weight storage unit 31 and an inference information storage unit 3.
2 and other storage means 33. The weighting storage means 31 learns in advance based on desired learning data in a target model to be inferred, specifies synapse weights between layers of the neural network, and stores synapse weights between the layers. is there.
The weighting storage means 31 corresponds to synapse storage means. The inference information storage unit 32 is a unit that stores necessary data when performing inference, and stores, for example, a value input to an input layer (first layer), the number of execution layers performing inference, and a final output value. are doing. The inference information storage means 32 also functions as a layer number storage means.

Next, regarding the operation of the inference apparatus of the present invention,
This will be described with reference to the flowchart of FIG. At the time of inference, an initial setting operation is first performed (step S1). The initial input control means 21 inputs and stores the value to be input to the input layer of the neural network in the input data storage means 13. This value is a value previously input from the input / output unit 2, or may be, for example, artificially input as a numerical value, or may be an output value of the sensor used in the model to be inferred. It is a value that is appropriately determined correspondingly.

Next, it is determined whether the number of execution layers of inference has reached the number of output layers (step S2). The number of execution layers is the number of layers that the one-layer neuron layer 12 is currently executing, and which layer the one-layer neuron layer 12 is sequentially shifted from the input layer (first layer) to the output layer (final layer). It is for monitoring whether the operation corresponding to is being performed.

If the number of execution layers is equal to or less than the maximum number of layers in step S2, the inference synchronization determination means 25 next determines whether or not it is the timing of weighting (step S3). At the weighting timing, the weight stored in the weight storage means 31 is set in the synapse 15 corresponding to the layer being executed by the weight control means 23. The synapse 15 has a weighting content that can be rewritten if necessary.

After the weighting, the output value of the further neuron layer 12 is stored in the output data storage means 14 (step S5). This means that one layer of operation has been executed, and the inference synchronization means 25 determines whether or not it is the timing for the data shift means 24 to perform a shift operation next (step S6). At the shift timing, the data shift means 24 shifts the content of the output data storage means 14 to the input data storage means 13, rewrites the content of the input data storage means 13 with the content of the output data storage means, and sets the content value. Output the corresponding value to the synapse. And
The number of execution layers monitored by the inference control means 26 is incremented by 1, and the operation of the next neuron layer is performed in step S4 described above.
The operations from to S8 are sequentially repeated. As a result, the operations of the hierarchical structure type neural network are sequentially executed one by one.

Although the number of layers storage means for storing the number of layers of the neural network is provided in the storage unit 30 in the above embodiment, the storage means may be an EEPROM in addition to the storage in ROM, RAM or the like. It would be even more convenient if the numbers could be electrically rewritten. Further, if the layer number designating means is provided in the input / output unit 2 and the like and the contents of the layer number storing means are rewritten, the size of the neural network can be changed more easily, which is more convenient.

Further, in the above embodiment, the contents of the weighting storage means 31 can be changed by exchanging the storage means such as ROM storing the various weights of the result learned in advance, or by the input / output unit 2. The weighting may be input and changed. Further, although not shown, the weighting may be determined by learning by the learning function.

If the number of execution layers exceeds the number of layers (maximum number of layers) corresponding to the output layer in step S2,
It is determined that the execution operation of the output layer is completed, and the contents of the current output data storage means 14 are used as the output of the inference operation corresponding to the neural network, and the contents are stored in the storage unit 30.
And ends the inference operation (step S9). Then, the content is displayed on a display unit (not shown) as necessary, or is output to another device as an inference result.

As another embodiment, the synapse 15 may be provided at a position between the neuron layer 12 and the output data storage means 14 and not at a position between the input data storage means 13. Further, the synapse 15 may be provided both between the input data storage means 13 and the one-layer neuron layer 12, and between the one-layer neuron layer 12 and the output data storage means 14.

[0026]

As described above, according to the present invention,
A single neuron layer, an input data storage means, and an output data storage means are provided, a synapse is provided between the single neuron layer and the input data storage means (output data storage means), and the weight of the synapse is set to the weight corresponding to the execution layer. However, the contents of the output data storage means are shifted to the input data storage means while being changed, so that there is an effect that the operation of the hierarchical neural network composed of a large number of layers in one neuron layer can be performed.

Further, since the shift operation is performed by the data shift means at the synchronization timing by the inference synchronization means, the operation of the hierarchical structure type neural network can be performed smoothly.

Further, in the present invention, since the maximum layer to be executed can be arbitrarily set, there is an effect that it is possible to cope with a neural network of an arbitrary size. Further, according to the present invention, even when it is applied to a large neural network, it can be realized only by taking a large storage capacity for weighting of synapses in the storage unit, and even a large-scale neural network can be realized in a small size.

[Brief description of drawings]

FIG. 1 is an explanatory diagram of a configuration of an inference device.

FIG. 2 is a flowchart showing an inference operation of an inference device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Inference device 2 Input / output unit 10 Execution unit 12 One layer neuron layer 13 Input data storage unit 14 Output data storage unit 15 Synapse 20 Control unit 21 Initial input control unit 23 Weighting control unit 24 Data shift unit 25 Inference synchronization unit 26 Inference control unit 30 storage unit 31 weighting storage unit 32 inference information storage unit

Claims (2)

[Claims] 1. A one-layer neuron layer that forms one layer of a plurality of neurons, a layer number storage unit that stores the number of layers of a neural network that performs inference, and a synapse weight corresponding to each layer of the neural network. Synapse storage means, input data storage means for storing data provided to the input side of the one-layer neuron layer, output data storage means for storing data on the output side of the one-layer neuron layer, and storage in the output data storage means Data shift means for shifting the stored contents to the input data storage means, and weights stored in the synapse storage means corresponding to the number of layers being executed are weighted to the synapses on the input / output side of the one neuron layer. And a weighting control means for performing the data until the number of layers stored in the layer number storage means is reached. The inference control in which the operation by the shift means and the weighting control means is repeated and the output on the output side of the one-layer neuron layer when the number of layers stored in the layer number storage means is reached becomes the inference output of the neural network An inference apparatus comprising means and. 2. A one-layer neuron layer that forms one layer of a plurality of neurons, a layer number storage unit that stores the number of layers of a neural network that performs inference, and a synapse weight corresponding to each layer of the neural network. Synapse storage means, input data storage means for storing data provided to the input side of the one-layer neuron layer, output data storage means for storing data on the output side of the one-layer neuron layer, and storage in the output data storage means Data shift means for shifting the stored contents to the input data storage means, and weights stored in the synapse storage means corresponding to the number of layers being executed are weighted to the synapses on the input / output side of the one neuron layer. Weighting control means for performing, shift timing by the data shift means, and the one-layer neuron layer Of the data based on the timing of the inference synchronization means until the number of layers stored in the layer number storage means is reached, and the inference synchronization means for timing the weighting of the input / output side synapses by the weighting control means. The inference control in which the operation by the shift means and the weight control means is repeated and the inference output of the neural network is based on the output on the output side of the one-layer neuron layer when the number of layers stored in the layer number storage means is reached An inference apparatus comprising means and.