Neural network apparatus

1. A neural network apparatus comprising:

a plurality of cores provided in accordance with a plurality of leaves in a tree structure, each core serving as a circuit that performs calculation or processing for part of elements of the neural network;

a plurality of routers provided in accordance with nodes other than the plurality of leaves in the tree structure;

a tree path that connects the plurality of cores and the plurality of routers such that data is transferred along the tree structure; and

a short-cut path that connects part of the plurality of routers such that data is transferred on a route differing from the tree path,

wherein the plurality of cores is configured to:

perform a normal calculation process in accordance with the neural network and a learning process for changing coefficients included in the neural network in parallel,

in the normal calculation process, propagate calculation data in a forward direction over a plurality of layers in the neural network, and

in the learning process, propagate learning data in a backward direction over the plurality of layers in the neural network, and

wherein the plurality of routers transmit data, which is output from each of the plurality of cores, to any of the plurality of cores serving as a transmission destination on one of routes included in the tree path and the short-cut path such that the calculation or the processing is performed in accordance with a structure of the neural network,

wherein among the plurality of routers, a first router connected to the short-cut path is configured to:

when received data is the calculation data, detect a transfer route for transmitting the received data to the core as the transmission destination from among the routes included in the tree path and the short-cut path, and transmit the received data, among the plurality of routers, to a next router that is next to the first router and that is determined in accordance with the transfer route, and

when the received data is the learning data, transmit the received data, among the plurality of routers, to a next router that is next to the first router and that is determined in accordance with the tree path.

2. The apparatus according to claim 1 , wherein the short-cut path includes a route connecting two or more routers, which are part of the plurality of routers, in a ring shape.

3. The apparatus according to claim 1 , wherein the short-cut path includes a route connecting two or more routers, which are part of the plurality of routers, in a line shape.

4. The apparatus according to claim 2 , wherein the short-cut path is a route connecting two or more routers disposed in a predetermined layer in the tree structure.

5. The apparatus according to claim 4 , wherein, when a number of layers at a leaf farthest from a root node is N, the short-cut path is a route connecting two or more routers disposed in an {N/2}-th, {(N+1)/2}-th, or {(N-1)/2}-th layer from the root node.

6. The apparatus according to claim 2 , wherein the short-cut path includes a route connecting two or more routers disposed in two different layers in the tree structure.

7. The apparatus according to claim 1 , wherein

the tree structure includes divisional areas,

each core disposed in each of the divisional areas is capable of transmitting data to another core disposed in a same divisional area without passing the root node, and

the short-cut path includes a route that connects two or more routers disposed in two or more different divisional areas in the divisional areas.

8. The apparatus according to claim 1 , further comprising substrates that have one-to-one correspondence with partial trees being divisions of the tree structure, each of the substrates being provided with two or more cores and two or more routers for two or more nodes and two or more leaves included in the corresponding partial tree,

wherein the short-cut path includes a route that connects the routers provided separately on two substrates of the substrates.

9. A data transfer method comprising:

transferring data of a neural network apparatus, wherein

the neural network apparatus comprises:

a plurality of cores provided in accordance with a plurality of leaves in a tree structure, each core serving as a circuit that performs calculation or processing for part of elements of the neural network;

a plurality of routers provided in accordance with nodes other than the plurality of leaves in the tree structure;

a tree path that connects the plurality of cores and the plurality of routers such that data is transferred along the tree structure; and

a short-cut path that connects part of the plurality of routers such that data is transferred on a route differing from the tree path, and

the method comprises:

by the plurality of cores, performing a normal calculation process in accordance with the neural network and a learning process for changing coefficients included in the neural network in parallel, wherein

in the normal calculation process, calculation data is propagated in a forward direction over a plurality of layers in the neural network, and

in the learning process, learning data is propagated in a backward direction over the plurality of layers in the neural network;

by the plurality of routers, transmitting data, which is output from each of the plurality of cores, to any of the plurality of cores serving as a transmission destination on one of routes included in the tree path and the short-cut path such that the calculation or the processing is performed in accordance with a structure of the neural network; and

by a first router connected to the short-cut path among the plurality of routers,

when received data is the calculation data, detecting a transfer route for transmitting the received data to the core as the transmission destination from among the routes included in the tree path and the short-cut path, and transmitting the received data, among the plurality of routers, to a next router that is next to the first router and that is determined in accordance with the transfer route, and

when the received data is the learning data, transmitting the received data, among the plurality of routers, to a next router that is next to the first router and that is determined in accordance with the tree path.