Claims:WE CLAIM:
1. A system (100) for improving classifications performed by an Artificial Neural Network (ANN) model, the system (100) comprising:
an ANN improvement device comprising at least one processor (101) and a computer-readable medium (102) storing instructions that, when executed by the at least one processor (101), cause the at least one processor (101) to perform operations comprising:
identifying (301), for a classification performed by the ANN model for an input, one or more activated neurons in one or more neural layers of the ANN model, wherein the one or more activated neurons are neurons achieving respective activation threshold values for the input and corresponding to one or more relevant features associated with the input;
analyzing (303) the one or more activated neurons in the one or more neural layers with respect to one or more Characteristic Feature Directive (CFDs) for the one or more neural layers and for a correct class of the input, wherein the one or more CFDs are generated after a training phase of the ANN model and based on one or more neurons in the one or more neural layers that are activated for a training input of the correct class;
determining (304) one or more differentiating neurons in the one or more neural layers based on the analysis, wherein the one or more differentiating neurons are neurons that are not activated as per the one or more CFDs for the correct class of the input; and
providing (305) one or more missing features based on the one or more differentiating neurons in the one or more neural layers for improving classifications performed by the ANN model.

2. The system (100) of claim 1, wherein providing (305) the one or more missing features further comprises providing the one or more missing features in a user-understandable format to a user.

3. The system (100) of claim 1, wherein operations further comprise re-training (306) the ANN model with new training data associated with the one or more missing features.

4. The system (100) of claim 1, wherein operations further comprise generating (302) the one or more CFDs for each neural layer of the ANN model and for each class of training data after the training phase of the ANN model.

5. The system (100) of claim 4, wherein generating (302) the one or more CFDs for a neural layer and for a class comprises:
identifying (307) a training input of the class in the training data;
identifying (308), for a classification performed by the ANN model for the training input of the class, one or more activated neurons in the neural layer, wherein the one or more activated neurons are neurons that achieve respective activation threshold values for the training input of the class; and
generating (309) the one or more CFDs for the neural layer and for the class based on the one or more activated neurons in the neural layer.

6. The system (100) of claim 1, wherein the one or more CFDs in a neural layer comprises a maximum value associated with the one or more activated neurons in the neural layer, a minimum value associated with the one or more activated neurons in the neural layer, a number of the one or more activated neurons in the neural layer, a list of the one or more activated neurons in the neural layer, and one or more mandatory neurons in the neural layer for performing the classification in the correct class.

7. The system (100) of claim 1, wherein the ANN model is a Convolutional Neural Network (CNN) model.

8. A method (300) of improving classifications performed by an Artificial Neural Network (ANN) model, the method comprising:
identifying (301), by an ANN improvement device and for a classification performed by the ANN model for an input, one or more activated neurons in one or more neural layers of the ANN model, wherein the one or more activated neurons are neurons achieving respective activation threshold values for the input and corresponding to one or more relevant features associated with the input;
analysing (303), by the ANN improvement device, the one or more activated neurons in the one or more neural layers with respect to one or more Characteristic Feature Directive (CFDs) for the one or more neural layers and for a correct class of the input, wherein the one or more CFDs are generated after a training phase of the ANN model and based on one or more neurons in the one or more neural layers that are activated for a training input of the correct class;
determining (304), by the ANN improvement device, one or more differentiating neurons in the one or more neural layers based on the analysis, wherein the one or more differentiating neurons are neurons that are not activated as per the one or more CFDs for the correct class of the input; and
providing (305), by the ANN improvement device, one or more missing features based on the one or more differentiating neurons in the one or more neural layers for improving classifications performed by the ANN model.

9. The method of claim 8, further comprising re-training (306) the ANN model with new training data associated with the one or more missing features.

10. The method of claim 8, further comprising generating (302) the one or more CFDs for each neural layer of the ANN model and for each class of training data after the training phase of the ANN model, wherein generating (302) the one or more CFDs for a neural layer and for a class comprises:
identifying (307) a training input of the class in the training data;
identifying (308), for a classification performed by the ANN model for the training input of the class, one or more activated neurons in the neural layer, wherein the one or more activated neurons are neurons that achieve respective activation threshold values for the training input of the class; and
generating (309) the one or more CFDs for the neural layer and for the class based on the one or more activated neurons in the neural layer.

Dated this 6th day of September, 2019

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