Optimization apparatus and method for controlling neural network
Arithmetic processing units individually compute values of a predetermined number of neurons amongst a plurality of neurons in a neural network, where the predetermined number of neurons belong to each arithmetic processing unit, and each of the predetermined number of neurons belonging to an arithmetic processing unit is connected to every other neuron belonging thereto. A control unit causes each arithmetic processing unit to perform the computation by providing dense-connection weight coefficients, which correspond to a dense-connection part of the evaluation function, associated with every connection of two neurons belonging thereto. The control unit calculates changes in energy using sparse-connection weight coefficients, which correspond to a sparse-connection part of the evaluation function, individually associated with a connection of two neurons belonging to different arithmetic processing units. Subsequently, the control unit causes each arithmetic processing unit to perform the computation by providing the changes.
1. An optimization apparatus for calculating a minimum of an evaluation function representing energy, the optimization apparatus comprising:
a plurality of arithmetic processing units configured to individually perform computation of values of a predetermined number of neurons amongst a plurality of neurons in a neural network, where the predetermined number of neurons belong to each of the plurality of arithmetic processing units, and each of the predetermined number of neurons belonging to one of the plurality of arithmetic processing units is connected to every other neuron belonging thereto; and
a control unit configured to:
cause each of the plurality of arithmetic processing units to perform the computation by providing the each of the plurality of arithmetic processing units with dense-connection weight coefficients associated with every connection of two neurons belonging thereto, the dense-connection weight coefficients corresponding to a dense-connection part of the evaluation function; and
calculate changes in the energy using sparse-connection weight coefficients individually associated with a connection of two neurons belonging to different arithmetic processing units, the sparse-connection weight coefficients corresponding to a sparse-connection part of the evaluation function, and subsequently cause each of the plurality of arithmetic processing units to perform the computation by providing the changes.
2. The optimization apparatus according to claim 1 , wherein:
the plurality of arithmetic processing units is connected one-on-one with the control unit.
3. The optimization apparatus according to claim 1 , wherein:
the plurality of arithmetic processing units is connected in a bus or ring topology, and one of the plurality of arithmetic processing units is connected to the control unit.
4. The optimization apparatus according to claim 1 , wherein:
the plurality of arithmetic processing units is connected in a two-dimensional mesh or torus topology, and one of the plurality of arithmetic processing units is connected to the control unit.
5. The optimization apparatus according to claim 4 , wherein:
each of the plurality of arithmetic processing units selects, based on indexes assigned to the dense-connection weight coefficients and the changes, one of a first arithmetic processing unit adjacent thereto in a first direction and a second arithmetic processing unit adjacent thereto in a second direction, through which the dense-connection weight coefficients and the changes are to be routed over a shorter communication path, and transmits the dense-connection weight coefficients and the changes to the one of the first arithmetic processing unit and the second arithmetic processing unit.
6. A method for controlling an optimization apparatus for calculating a minimum of an evaluation function representing energy, the method comprising:
performing, by each of a plurality of arithmetic processing units, computation of values of a predetermined number of neurons amongst a plurality of neurons in a neural network, where the predetermined number of neurons belong to each of the plurality of arithmetic processing units, and each of the predetermined number of neurons belonging to one of the plurality of arithmetic processing units is connected to every other neuron belonging thereto; and
causing, by a control unit, each of the plurality of arithmetic processing units to perform the computation by providing the each of the plurality of arithmetic processing units with dense-connection weight coefficients associated with every connection of two neurons belonging thereto, the dense-connection weight coefficients corresponding to a dense-connection part of the evaluation function, calculating changes in the energy using sparse-connection weight coefficients individually associated with a connection of two neurons belonging to different arithmetic processing units, the sparse-connection weight coefficients corresponding to a sparse-connection part of the evaluation function, and subsequently causing each of the plurality of arithmetic processing units to perform the computation by providing the changes.