Small-world nets for fast neural network training and execution
A method, a system, and a computer program product for fast training and/or execution of neural networks. A description of a neural network architecture is received. Based on the received description, a graph representation of the neural network architecture is generated. The graph representation includes one or more nodes connected by one or more connections. At least one connection is modified. Based on the generated graph representation, a new graph representation is generated using the modified at least one connection. The new graph representation has a small-world property. The new graph representation is transformed into a new neural network architecture.
1. A computer-implemented method, comprising:
receiving, by one or more processors, a description of a neural network architecture;
generating, by the one or more processors, based on the description, a graph representation of the neural network architecture, the graph representation comprising one or more nodes connected by one or more connections;
modifying, by the one or more processors, at least one connection in the one or more connections to optimize a topology of the neural network architecture as identified through an iterative rewiring process;
generating, by the one or more processors, based on the graph representation, a new graph representation using the at least one connection, wherein the new graph representation has a small-world property; and
transforming, by the one or more processors, the new graph representation into a new neural network architecture possessing a minimum mixing time.
2. The method of claim 1 , wherein the modifying the at least one connection is executed based on a predetermined probability, wherein the predetermined probability is selected in an interval between 0 and 1.
3. The method according to of claim 2 , further comprising
modifying each connection in the one or more connections; and
generating the new graph representation having a maximum small-world property selected from a plurality of small-world properties determined for each new graph representation based on a plurality of probabilities in the interval between 0 and 1.
4. The method of claim 1 , wherein the description of the neural network architecture comprises a plurality of layers having a plurality of neurons, wherein each neuron corresponds to a node in the graph representation and the one or more connections comprise connections between one or more layers in the plurality of layers formed in a small-world graph representation.
5. The method of claim 3 , wherein
the at least one connection is between a first input node and a first output node in a plurality of nodes comprising the one or more nodes; and
the at least one connection is between the first input node and a second output node in the plurality of nodes, wherein the at least one connection is selected using the predetermined probability and one or more constraints.
6. The method according to of claim 5 , wherein the at least one connection is longer than the at least one connection.
7. The method of claim 3 , wherein the new neural network architecture corresponds to a small-world neural network.
8. The method of claim 1 , wherein a total number of the one or more connections in the graph representation is equal to a total number of one or more connections in the new graph representation.
9. The method according to claim 1 , wherein an application programming interface is configured to perform at least one of receiving the description of the neural network architecture, generating of the graph representation, modifying, the generating of the new graph representation, and transforming to the new neural network architecture.
10. A system comprising:
at least one programmable processor; and
a non-transitory machine-readable medium storing instructions that, when executed by the at least one programmable processor, cause the at least one programmable processor to perform operations comprising:
receiving a description of a neural network architecture;
generating, based on the description, a graph representation of the neural network architecture, the graph representation including one or more nodes connected by one or more connections;
modifying at least one connection in the one or more connections to optimize a topology of the neural network architecture as identified through an iterative rewiring process;
generating, based on the graph representation, a new graph representation using the at least one connection, wherein the new graph representation has a small-world property; and
transforming the new graph representation into a new neural network architecture possessing a minimum mixing time.
11. The system of claim 10 , wherein the modifying the at least one connection is executed based on a predetermined probability, wherein the predetermined probability is selected in an interval between 0 and 1.
12. The system of claim 11 , wherein the operations further comprise
modifying each connection in the one or more connections; and
generating the new graph representation having a maximum small-world property selected from a plurality of small-world properties determined for each new graph representation based on a plurality of probabilities in the interval between 0 and 1.
13. The system of claim 10 , wherein the description of the neural network architecture includes a plurality of layers having a plurality of neurons, wherein each neuron corresponds to a node in the graph representation and the one or more connections include connections between one or more layers in the plurality of layers formed in a small-world graph representation.
14. The system of claim 12 , wherein
the at least one connection is between a first input node and a first output node in a plurality of nodes comprising the one or more nodes; and
the at least one connection is between the first input node and a second output node in the plurality of nodes, wherein the at least one connection is selected using the predetermined probability and one or more constraints.
15. The system of claim 14 , wherein the modified at least one connection is longer than the at least one connection.
16. The system of claim 12 , wherein the new neural network architecture corresponds to a small-world neural network.
17. The system of claim 10 , wherein a total number of the one or more connections in the graph representation is equal to a total number of one or more connections in the new graph representation.
18. The system of claim 10 , wherein an application programming interface is configured to perform at least one of receiving description of the neural network architecture, generating of the graph representation, modifying, the generating of the new graph representation, and transforming to the new neural network architecture.
19. A computer program product comprising a non-transitory machine-readable medium storing instructions that, when executed by at least one programmable processor, cause the at least one programmable processor to perform operations comprising:
receiving a description of a neural network architecture;
generating, based on the description, a graph representation of the neural network architecture, the graph representation including one or more nodes connected by one or more connections;
modifying at least one connection in the one or more connections to optimize a topology of the neural network architecture as identified through an iterative rewiring process;
generating, based on the graph representation, a new graph representation using the at least one connection, wherein the new graph representation has a small-world property; and
transforming the new graph representation into a new neural network architecture possessing a minimum mixing time.
20. The computer program product of claim 19 , wherein the modifying the at least one connection is executed based on a predetermined probability, wherein the predetermined probability is selected in an interval between 0 and 1.