JPH0594429A - Boltzmann machine simulation device - Google Patents

Abstract

PURPOSE:To provide a simulation whose efficiency is satisfactory in which plural processors can be asynchronously operated by utilizing the asynchronous operation of a Boltsmann machine. CONSTITUTION:The output values of the entire nodes of a network are stored in a node vector 4, and the weight of a link to be inputted to a node group included in a partial network charged by a processor 1 is stored in a partial weight matrix 5. An arithmetic control unit 3 calculates the output of the node of the partial network from both the node vector 4 and the partial weight matrix 5. A communication control unit 6 reloads the values of the elements of the node vectors 4 of the other processors 1 into the output value calculated by the arithmetic control unit 3 through a between-processor communication path 2 based on the identification number of the node. At that time, the other processors 1 whose node vectors 4 are reloaded continue the same simulation asynchronously with the main processor 1.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a simulation apparatus for a Boltzmann machine which is a neural network model.

[0002]

2. Description of the Related Art Conventionally, when a Boltzmann machine simulation is processed in parallel by a plurality of processors, in order to transfer the value of a node processed by one processor to another processor, the plurality of processors synchronously transfer data. Was there.

[0003]

However, there is a drawback that synchronization of a plurality of processors becomes an overhead in parallel processing, which hinders improvement in processing speed due to parallel processing.

An object of the present invention is to provide a high-speed Boltzmann machine simulation by utilizing a characteristic of the Boltzmann machine, which is basically an asynchronous calculation model, and operating a plurality of processors asynchronously.

[0005]

The present invention is a Boltzmann machine simulation apparatus, which is a neural network model in which each node operates asynchronously and stochastically, and has (a) a plurality of (M) processors. (B) has an inter-processor communication path for transferring data between the plurality of processors, and (c) represents outputs of a plurality of (N) nodes of a network to be simulated as a one-dimensional node vector, Multiple networks (N x
N) weighted links are represented as a N × N two-dimensional weight matrix, (d) the network is divided into M sub-networks, each of which is assigned to the M processors, and (e) The processor has a node vector that holds the outputs of all the nodes of the network, and a partial weight matrix that holds the weights of the links that become the inputs to the nodes included in the partial network that the processor is in charge of processing, (F) The processor further includes an arithmetic unit for simulating the partial network, which is processed by the processor, from the node vector and the partial weight matrix, and the processor of another processor via the interprocessor communication path. It has a communication control unit that rewrites the value of the node vector asynchronously with other processors,
(G) The processor calculates the output of a node of the sub-network in charge of processing asynchronously with the other processors, and the node vector held by the other processors via the inter-processor communication path. Is rewritten asynchronously with the other processor using the communication control unit.

[0006]

In the present invention, the processor has the node vector holding the outputs of all the nodes of the network to be simulated, and becomes the input to the node group included in the partial network in which the processor is in charge of processing. Since we have a partial weight matrix that holds the link weights,
The output of a node included in the sub-network in charge of the processor can be calculated asynchronously with other processors. The calculation result of the output of the node needs to be transferred to the node vector held by another processor, but since the Boltzmann machine is an asynchronous calculation model in which each node changes the output asynchronously, the output of the node is Can be performed asynchronously with the progress of the simulation of the processor, the processor having the communication control unit for rewriting the value of the node vector of the other processor asynchronously with the other processor, Data transfer is realized by this. In this way, we have realized a parallel simulation of a Boltzmann machine in which multiple processors can proceed asynchronously.

[0007]

Embodiments of the present invention will now be described with reference to the drawings. FIG. 1 is an overall configuration diagram of an embodiment of the present invention, which includes a plurality of processors 1 and an interprocessor communication path 2.

The processor 1 includes an arithmetic control unit 3 and
It is composed of a node vector 4, a partial weight matrix 5, and a communication control unit 6, and the node vector 4 stores output values of all nodes of the network to be simulated. Further, the partial weight matrix 5 stores the weight of the link which is the input of the node group included in the partial network, which the processor 1 is in charge of. Arithmetic control unit 3
Calculates the output of the node included in the partial network which this processor is in charge of from the node vector 4 and the partial weight matrix 5. This value must be transferred to the node vector 4 held by another processor. The arithmetic and control unit 3 notifies the communication control unit 6 of the node identification number (node address) to be transferred and the calculated output value.
The communication control unit 6 calculates the value of the element designated by the node identification number among the elements of the node vector 4 of another processor 1 via the interprocessor communication path 2 based on the node identification number. The output value. At this time, the other processor 1 in which the node vector 4 is rewritten continues the same simulation asynchronously with the main processor 1. Since the Boltzmann machine is a neural network model in which each node can operate asynchronously, even if another processor 1 simultaneously refers to a node rewritten by this processor 1, the simulation does not become inaccurate.

As described above, each processor 1 can perform a simulation without synchronizing with another processor 1, and a high-speed simulation without the overhead of synchronization between processors can be realized.

[0010]

As described above, in the present invention, the processor has the node vector holding the output of all the nodes of the network to be simulated, and is included in the partial network in which the processor is in charge of processing. Since the partial weight matrix that holds the weight of the link that is the input to the node group is included, the output of the node included in the partial network in charge of the processor can be calculated asynchronously with the other processors. The calculation result of the output of the node needs to be transferred to the node vector held by another processor, but since the Boltzmann machine is an asynchronous calculation model in which each node changes the output asynchronously, the output of the node is Can be performed asynchronously with the progress of the simulation of the processor, the processor having the communication control unit for rewriting the value of the node vector of the other processor asynchronously with the other processor, Data transfer is realized by this. In this way, we have realized a parallel simulation of a Boltzmann machine in which multiple processors can proceed asynchronously.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention.

[Explanation of symbols]

 1 processor 2 communication path between processors 3 arithmetic control unit 4 node vector 5 partial weight matrix 6 communication control unit

Claims (1)

[Claims] 1. A device for simulating a Boltzmann machine, which is a neural network model in which each node operates asynchronously and stochastically, comprising: (a) a plurality of (M) processors, and (b) between the plurality of processors. And has a communication path between processors for transferring data, and (c) a plurality of (N) networks to be simulated.
The output of each node is represented as a one-dimensional node vector, and a plurality of (N × N) weighted links of the network are represented as an N × N two-dimensional weight matrix, and (d) the network is It is divided into M sub-networks and assigned to each of the M processors, (e) the processor holds a node vector holding outputs of all nodes of the network, and the sub-network in which the processor is in charge of processing. A partial weight matrix for holding the weight of the link that is an input to the node group included in (f), the processor further including the node vector and the partial weight matrix, An arithmetic unit for simulating a network and a value of the node vector of another processor via the interprocessor communication path And a communication control unit that rewrites asynchronously with the other processor, and (g) the processor calculates the output of a node of the partial network in charge of processing asynchronously with the other processor, and calculates it between the processors. A Boltzmann machine simulation apparatus characterized in that the node vector held by another processor is rewritten asynchronously with another processor via the communication path by using the communication control unit.