US 12,353,509 B1
System, method, and computer program for identifying a hidden class
Eyal Felstaine, Herzliya (IL); Gad Yitzhak Weissman, Hod Hasharon (IL); Marina Ankri, Karnei Shomron (IL); and Nimrod Sandlerman, Ramat Gan (IL)
Assigned to AMDOCS DEVELOPMENT LIMITED, Limassol (CY)
Filed by Amdocs Development Limited, Limassol (CY)
Filed on Dec. 16, 2020, as Appl. No. 17/124,331.
Int. Cl. G06F 18/214 (2023.01); G06F 18/24 (2023.01); G06N 20/00 (2019.01); G06V 30/242 (2022.01)
CPC G06F 18/2155 (2023.01) [G06F 18/24 (2023.01); G06N 20/00 (2019.01); G06V 30/242 (2022.01)] 9 Claims
OG exemplary drawing
 
1. A method for creating a new class of instances, the method comprising:
obtaining a first dataset comprising a plurality of instances;
using a first AI model to classify the plurality of instances of the first dataset into a plurality of classes;
selecting a class of the plurality of classes and executing a respective verification action on instances of the selected class to generate for each of the instances of the selected class a result that is selected from a group consisting of:
a positive result indicating that the instance is part of the selected class, and
a lack of the positive result indicating that the instance is not part of the selected class;
collecting a first subset of instances of the selected class for which the respective verification action resulted in the lack of the positive result;
identifying the first subset of instances as having less than a threshold number of instances required for training the first AI model to be able to recognize instances in the first subset as being part of a particular class;
synthesizing a plurality of artificial instances by modifying at least one instance in the first subset of instances for which the respective action resulted in the lack of the positive result, wherein the modification:
changes at least one parameter of the at least one instance away from a characteristic value of a second subset of instances of the selected class for which the respective action resulted in the positive result, or
changes at least one parameter of the at least one instance toward a characteristic value of the first subset of instances for which the respective action resulted in the lack of the positive result; and
combining the plurality of artificial instances and the first subset of instances into a second dataset; and
training a second AI model to recognize instances in the second dataset as being part of the particular class.