Claims:WE CLAIM:
1. A method of improving performance of an artificial neural network (ANN), the method comprising:
receiving, by an ANN improvement device, a weight matrix comprising an original weight of each neural node in each layer of the ANN;
for each unique combination of at least two neural nodes in each layer,
determining, by the ANN improvement device, a relative advantage value for one of the at least two neural nodes in a given layer with respect to remaining of the at least two neural nodes in the given layer based on actual inputs and standard inputs to the at least two neural nodes; and
determining, by the ANN improvement device, a modified weight of each of the at least two neural nodes based on the relative advantage value;
executing, by the ANN improvement device, an elimination decision for each neural node in each layer based on a corresponding final modified weight, wherein the final modified weight of a given neural node in a given layer is the modified weight of the given neural node upon exhausting each unique combination of the at least two neural nodes in the given layer; and
updating, by the ANN improvement device, the weight matrix based on the final modified weight of each remaining neural node in each layer.
2. The method of claim 1, further comprising generating the weight matrix by building and training the ANN for a target application.
3. The method of claim 1, wherein determining the relative advantage value for the one of the at least two neural nodes in the given layer with respect to the remaining of the at least two neural nodes in the given layer comprises:
determining actual outputs and standard outputs of the at least two neural nodes for the actual inputs and standard inputs respectively based on the original weights or the modified weights of the at least two neural nodes;
determining a relative advantage of the one of the at least two neural nodes with respect to the remaining of the at least two neural nodes by comparing the actual outputs and the standard outputs, each normalized with respect to one of the standard outputs or the actual outputs; and
determining the relative advantage value based on a difference between the actual outputs, normalized with respect to the one of the standard outputs or the actual outputs.
4. The method of claim 3, wherein the at least two neural nodes comprises two neural nodes, and wherein the relative advantage value comprises an average of the difference between the actual outputs, normalized with respect to the one of the standard outputs or the actual outputs.
5. The method of claim 1, wherein determining the modified weight of each of the at least two neural nodes comprises:
increasing an original or a previously modified weight of the one of the at least two neural nodes by a first value proportional to the relative advantage value; and
decreasing an original or a previously modified of each of the remaining of the at least two neural nodes by a second value proportional to the relative advantage value.
6. The method of claim 1, wherein executing the elimination decision comprises removing a given neural node for which the final modified weight is about zero or about same as the original weight.
7. The method of claim 1, wherein updating the weight matrix comprises replacing the original weight of each remaining neural node in each layer with the corresponding final modified weight.
8. A system for improving performance of an artificial neural network (ANN), the system comprising:
an ANN improvement device comprising at least one processor and a computer-readable medium storing instructions that, when executed by the at least one processor, cause the at least one processor to perform operations comprising:
receiving a weight matrix comprising an original weight of each neural node in each layer of the ANN;
for each unique combination of at least two neural nodes in each layer,
determining a relative advantage value for one of the at least two neural nodes in a given layer with respect to remaining of the at least two neural nodes in the given layer based on actual inputs and standard inputs to the at least two neural nodes; and
determining a modified weight of each of the at least two neural nodes based on the relative advantage value;
executing an elimination decision for each neural node in each layer based on a corresponding final modified weight, wherein the final modified weight of a given neural node in a given layer is the modified weight of the given neural node upon exhausting the each unique combination of at least two neural nodes in the given layer; and
updating the weight matrix based on the final modified weight of each remaining neural node in each layer.
9. The system of claim 8, wherein determining the relative advantage value for the one of the at least two neural nodes in the given layer with respect to the remaining of the at least two neural nodes in the given layer comprises:
determining actual outputs and standard outputs of the at least two neural nodes for the actual inputs and standard inputs respectively based on the original weights or the modified weights of the at least two neural nodes;
determining a relative advantage of the one of the at least two neural nodes with respect to the remaining of the at least two neural nodes by comparing the actual outputs and the standard outputs, each normalized with respect to one of the standard outputs or the actual outputs; and
determining the relative advantage value based on a difference between the actual outputs, normalized with respect to the one of the standard outputs or the actual outputs.
10. The system of claim 8, wherein determining the modified weight of each of the at least two neural nodes comprises:
increasing an original or a previously modified weight of the one of the at least two neural nodes by a first value proportional to the relative advantage value; and
decreasing an original or a previously modified of each of the remaining of the at least two neural nodes by a second value proportional to the relative advantage value.

Dated this 27th day of December, 2018

Madhusudan S.T.
Of K&S Partners
Agent for the Applicant
IN/PA-1297
, Description:TECHNICAL FIELD
This disclosure relates generally to artificial neural network (ANN), and more particularly to method and system for improving performance of an ANN.