US 12,293,266 B2
Learning data augmentation policies
Vijay Vasudevan, Los Altos Hills, CA (US); Barret Zoph, San Francisco, CA (US); Ekin Dogus Cubuk, Sunnyvale, CA (US); and Quoc V. Le, Sunnyvale, CA (US)
Assigned to Google LLC, Mountain View, CA (US)
Filed by Google LLC, Mountain View, CA (US)
Filed on Feb. 22, 2024, as Appl. No. 18/584,625.
Application 18/584,625 is a continuation of application No. 17/061,103, filed on Oct. 1, 2020, granted, now 12,033,038.
Application 17/061,103 is a continuation of application No. 16/417,133, filed on May 20, 2019, granted, now 10,817,805, issued on Oct. 27, 2020.
Claims priority of provisional application 62/673,777, filed on May 18, 2018.
Prior Publication US 2024/0242125 A1, Jul. 18, 2024
Int. Cl. G06N 20/00 (2019.01)
CPC G06N 20/00 (2019.01) 20 Claims
OG exemplary drawing
 
1. A method comprising, at each of multiple time steps:
generating, based on quality measures of data augmentation policies generated at one or more previous time steps, a current data augmentation policy that defines a procedure for transforming training inputs for a neural network, wherein:
for each previous time step, the quality measure of the data augmentation policy generated at the previous time step characterizes a performance measure of the neural network after the neural network has been trained using the data augmentation policy generated at the previous time step; and
the current data augmentation policy defines the procedure for transforming training inputs by defining, for each transformation operation in a sequence of transformation operations for the current data augmentation policy:
(i) a respective probability distribution over a space of transformation types for the transformation operation, and
(ii) a respective probability distribution over a space of transformation magnitudes for the transformation operation;
training the neural network using the current data augmentation policy, comprising:
selecting a batch of training inputs;
determining an augmented batch of training inputs by transforming the training inputs in the batch of training inputs in accordance with the current data augmentation policy, comprising, for each training input in the batch of training inputs:
for each transformation operation in the sequence of transformation operations for the current data augmentation policy:
selecting a transformation type for the transformation operation in accordance with the probability distribution over the space of transformation types for the transformation operation;
selecting a transformation magnitude for the transformation operation in accordance with the probability distribution over the space of transformation magnitudes for the transformation operation;
transforming the training input by applying the transformation operation with the selected transformation type and transformation magnitude; and
adjusting current values of parameters of the neural network by training the neural network on the augmented batch of training inputs that have been transformed in accordance with the current data augmentation policy by a machine learning training technique to optimize an objective function; and
determining a quality measure of the current data augmentation policy using the neural network after the neural network has been trained using the current data augmentation policy.