Calculation processing system, program creation method, and program creation program
원문보기
IPC분류정보
국가/구분
United States(US) Patent
등록
국제특허분류(IPC7판)
G06N-003/08
G06N-003/12
출원번호
US-0125540
(2009-10-23)
등록번호
US-8984103
(2015-03-17)
우선권정보
JP-2008-274282 (2008-10-24)
국제출원번호
PCT/JP2009/068271
(2009-10-23)
§371/§102 date
20110421
(20110421)
국제공개번호
WO2010/047388
(2010-04-29)
발명자
/ 주소
Suzuki, Hideaki
Sawai, Hidefumi
Ohsaki, Hiroyuki
출원인 / 주소
National Institute of Information and Communications Technology (NIICT)
대리인 / 주소
McDermott Will & Emery LLP
인용정보
피인용 횟수 :
0인용 특허 :
7
초록▼
A calculation processing apparatus includes a monitor, a CPU, a memory and a hard disk. The hard disk stores an initial program input from outside, a network creation program, a network modifying program, network information, node operation definition, and learning variables. The CPU executes the ne
A calculation processing apparatus includes a monitor, a CPU, a memory and a hard disk. The hard disk stores an initial program input from outside, a network creation program, a network modifying program, network information, node operation definition, and learning variables. The CPU executes the network creation program, and creates network information related to a network representing an algorithm structure of the initial program. Further, the CPU executes the network modifying program and modifies the network information based on the result of calculation by the network, using a learning algorithm.
대표청구항▼
1. A calculation processing system, comprising: an input unit receiving, from outside of the calculation processing system, programming code for a program including an algorithm;a network creation unit creating, based on the programming code for the program received by the input unit, a network to h
1. A calculation processing system, comprising: an input unit receiving, from outside of the calculation processing system, programming code for a program including an algorithm;a network creation unit creating, based on the programming code for the program received by the input unit, a network to have a topology representing an algorithm structure of the algorithm of the program, wherein the network created by the network creation unit includes a plurality of operation nodes and a plurality of edges each connecting two of said operation nodes according to the topology of the network, andwherein said plurality of operation nodes include operation nodes corresponding to various operations of the program and an operation node performing calculation on a regulating value representing an algorithm flow;a calculation executing unit executing a calculation using the network representing the algorithm structure of the algorithm of the program; anda network updating unit for changing, using a result of calculation by said calculation executing unit and a learning algorithm, the topology of the network without changing said result of calculation,wherein the network creation unit determines the topology representing the algorithm structure of the algorithm of the program received by the input unit, and creates the network to have the determined topology by adding or deleting an operation node and configuring edges connecting the operation node, andwherein the network creation unit converts the program to a network in ATN (Algorithmically Transitive Network) format. 2. The calculation processing system according to claim 1, wherein said plurality of operation nodes include an operation node having a node variable; andsaid network updating unit calculates an error of said result of calculation from a teaching value, calculates contribution of said node variable to said error, and modifies said node variable based on said contribution. 3. The calculation processing system according to claim 1, wherein said plurality of operation nodes include operation nodes corresponding to four arithmetic operations. 4. The calculation processing system according to claim 2, wherein said network updating unit performs a process of changing, adding or deleting said operation node, on said network. 5. The calculation processing system according to claim 4, wherein said network updating unit changes said operation node, of which output value is kept unchanged for a prescribed number of times continuously, to a constant node outputting a constant value. 6. The calculation processing system according to claim 4, wherein said network updating unit selects at random two or more of said operation nodes from said plurality of operation nodes, from nearby nodes, and creates a new said operation node to be connected to said selected operation nodes. 7. The calculation processing system according to claim 4, wherein said network updating unit selects at random a first operation node and a second operation node from said plurality of operation nodes, from nearby nodes, and creates a bridge from said second operation node to said first operation node. 8. The calculation processing system according to claim 4, wherein said network updating unit selects at random said operation node from said plurality of operation nodes, from nearby nodes, and creates a branch of said selected node. 9. The calculation processing system according to claim 4, wherein said network updating unit rewrites a plurality of constant nodes each outputting a constant value, to one said constant node in accordance with coupling rule. 10. The calculation processing system according to claim 4, wherein said network updating unit divides said operation node having said node variable modified successively for a prescribed number of times or more. 11. The calculation processing system according to claim 2, wherein said network updating unit calculates an energy error based on said error, calculates differential coefficient of said energy error with respect to said node variable as said contribution, and updates said variable by subtracting a product of said differential coefficient and a learning coefficient, from said variable. 12. The calculation processing system according to claim 11, wherein said learning coefficient is a value determined such that said energy error is multiplied (1−η1p) times (where η1pis a real number larger than 0 and smaller than 1), by the update of said variable. 13. The calculation processing system according to claim 1, wherein said program received by said input unit is source code described in a high level language; andsaid network creation unit converts said source code to a language-independent intermediate code, and creates said network based on said intermediate code. 14. The calculation processing system according to claim 1, comprising a plurality of calculation processing apparatuses operating in a distributed type calculation processing environment; whereineach of said calculation processing apparatuses processes the process executed by said calculation executing unit and the process performed by said network updating unit, in a distributed manner. 15. A method of creating a program using a calculation processing system, comprising steps of: receiving, in said calculation processing system from outside of the calculation processing system, programming code for a program including an algorithm;creating, in said calculation processing system, based on the received programming code for the program, a network to have a topology representing an algorithm structure of the algorithm of the program, wherein the created network includes a plurality of operation nodes and a plurality of edges each connecting said operation nodes according to the topology of the network, andwherein said plurality of operation nodes include operation nodes corresponding to various operations of the program and an operation node performing calculation on a regulating value representing an algorithm flow;executing, in said calculation processing system, a calculation using the created network representing the algorithm structure of the algorithm of the program; andmodifying, in said calculation processing system, using a result of said calculation and a learning algorithm, the topology of the network while maintaining said result of calculation,wherein the step for creating the network includes adding or deleting an operation node and configuring edges connecting the operation node to create the network to have the topology representing the algorithm structure of the received program, andwherein the creating the network comprises converting the program to a network in ATN (Algorithmically Transitive Network) format. 16. A non-transitory computer readable recording medium storing a program creation program causing a calculation processing system to create a program, wherein said program creation program causes the computer to execute a program creation program including steps of:causing said calculation processing system to create, based on programming code received from outside of the calculating processing system for a program including an algorithm, a network to have a topology representing an algorithm structure of the algorithm of the program, wherein the created network includes a plurality of operation nodes and a plurality of edges each connecting said operation nodes according to the topology of the network, andwherein said plurality of operation nodes include operation nodes corresponding to various operations of the program and an operation node performing calculation on a regulating value representing an algorithm flow;causing said calculation processing system to execute a calculation using the network representing the algorithm structure of the algorithm of the program; andcausing said calculation processing system to modify, using a result of said calculation and a learning algorithm, the topology of said network while maintaining said result of calculation,wherein the calculation processing system determines the topology representing the algorithm structure of the algorithm of the received program, and creates the network to have the determined topology by adding or deleting an operation node and configuring edges connecting the operation node, andwherein the creating the network comprises converting the program to a network in ATN (Algorithmically Transitive Network) format. 17. A non-transitory computer readable recording medium storing a program creation program causing a calculation processing system to create a program, the program creation program comprising steps of: causing said calculation processing system to create, based on programming code received from outside of the calculating processing system for a program including an algorithm, a network to have a topology representing an algorithm structure of the algorithm of the program, wherein the created network includes a plurality of operation nodes and a plurality of edges each connecting said operation nodes according to the topology of the network, andwherein said plurality of operation nodes include operation nodes corresponding to various operations of the program and an operation node performing calculation on a regulating value representing an algorithm flow;causing said calculation processing system to execute a calculation using the network representing the algorithm structure of the algorithm of the program; andcausing said calculation processing system to modify, using a result of said calculation and a learning algorithm, the topology of said network while maintaining said result of calculation,wherein the calculation processing system determines the topology representing the algorithm structure of the algorithm of the received program, and creates the network to have the determined topology by adding or deleting an operation node and configuring edges connecting the operation node, andwherein the creating the network comprises converting the program to a network in ATN (Algorithmically Transitive Network) format. 18. The calculation processing system according to claim 1, wherein the network creation unit associates with operation nodes of the plurality of operation nodes respective computational operations to be performed by the operation nodes on inputs to the operation nodes to compute an output of each respective operation node according to the computation operation associated with the operation node for transfer along an edge to another operation node.
연구과제 타임라인
LOADING...
LOADING...
LOADING...
LOADING...
LOADING...
이 특허에 인용된 특허 (7)
Tang,Chunqiang; Chang,Rong Nickle; Ward,Christopher, Apparatus, system, and method for reliable, fast, and scalable multicast message delivery in service overlay networks.
Furuta Toshiyuki (Yokohama JPX) Kitaguchi Takashi (Yokohama JPX) Eguchi Hirotoshi (Yokohama JPX), Signal processing apparatus having at least one neural network.
※ AI-Helper는 부적절한 답변을 할 수 있습니다.