| CPC G06N 3/08 (2013.01) [G06N 5/022 (2013.01); G06N 3/045 (2023.01); G06N 7/01 (2023.01)] | 18 Claims |
|
1. An apparatus for scheduling a data augmentation technique, the apparatus comprising:
at least one memory configured to store thereon computer program code; and
at least one processor configured to access the at least one memory and operate according to the computer program code, wherein the computer program code causes the at least one processor to perform:
generating a reduced data set by extracting data from a given training data set based on a data distribution of the given training data set;
training each of a plurality of first neural network models, reduced from a main model, based on a data augmentation technique for each of a given plurality of operations and the reduced data set;
generating a reduced operation pool by extracting operations from the given plurality of operations based on each respective performance of each of the trained plurality of first neural network models;
training each of a plurality of second neural network models, each of which having the same network structure as a corresponding first neural network model among the plurality of first neural network models, based on a data augmentation technique for each operation included in the reduced operation pool and the reduced data set, and performing iterative training comprising multiple preset units of training while updating, for each preset unit of training, the data augmentation technique for each operation included in the reduced operation pool based on a training result of each of the plurality of second neural network models; and
determining an update schedule, from a plurality of update schedules respectively corresponding to updates of the data augmentation technique for the multiple preset units of training, as an optimal schedule for the main model based on a final training result of each of the plurality of second neural network models.
|