Claims:We Claim
1. A method of minimizing impact of a faulty node associated with an artificial neural network, the method comprising:
detecting, by a rerouting system, a faulty node associated with an artificial neural network based on testing the artificial neural network periodically with a training dataset, wherein the faulty node causes a faulty path in the artificial neural network;
identifying, by the rerouting system, a plurality of alternate paths to reroute the faulty path, wherein the plurality of alternate paths is identified based on at least one network connection between one or more upstream nodes of the faulty node and one or more downstream nodes of the faulty node; and
rerouting, by the rerouting system, the faulty path by assigning one or more weights associated with the faulty node to one or more nodes associated with the plurality of alternate paths.
2. The method as claimed in claim 1, wherein the faulty node is detected if a predefined response to the training dataset is not received.
3. The method as claimed in claim 1, wherein the faulty path is rerouted if the faulty path is associated with one or more priority nodes.
4. The method as claimed in claim 3, wherein the one or more priority nodes are identified based on a relevance heat map associated with the training dataset, wherein the relevance heat map is generated through at least one of Layer-wise Relevance Propagation (LRP) and Sensitivity Analysis (SA).
5. The method as claimed in claim 4, wherein the faulty path is rerouted based on an impact of the one or more priority nodes on classification associated with the artificial neural network.
6. The method as claimed in claim 5, wherein the impact of the one or more priority nodes is determined by:
setting, by the rerouting system, one or more weights associated with each priority node to zero successively; and
recording, by the rerouting system, a change in classification associated with each priority node based on setting the one or more priority nodes to zero, wherein the change in classification relates to the impact of each priority node of the one or more priority nodes.
7. The method as claimed in claim 6, further comprises selecting an alternate path from the plurality of alternate paths to rectify the change in classification and restore the relevance heat map associated with the training dataset.
8. The method as claimed in claim 1, wherein the faulty node is associated with an artificial intelligence (AI) hardware accelerator.
9. A system for minimizing impact of a faulty node associated with an artificial neural network, the system comprising:
a processor; and
a memory communicatively coupled to the processor, wherein the memory stores processor executable instructions, which on execution causes the processor to:
detect a faulty node associated with an artificial neural network based on testing the artificial neural network periodically with a training dataset, wherein the faulty node causes a faulty path in the artificial neural network;
identify a plurality of alternate paths to reroute the faulty path, wherein the plurality of alternate paths is identified based on at least one network connection between one or more upstream nodes of the faulty node and one or more downstream nodes of the faulty node; and
reroute the faulty path by assigning one or more weights associated with the faulty node to one or more nodes associated with the plurality of alternate paths.
10. The system as claimed in claim 9, wherein the processor is configured to:
detect the faulty node if a predefined response to the training dataset is not received;
rereoute the faulty path if the faulty path is associated with one or more priority nodes.
, Description:Technical Field
[001] This disclosure relates to the field of artificial neural network and more particularly to a method and system for rerouting around a faulty node in an artificial neural network.