Abstract:
"The performance of modern Convolutional Neural Networks (CNN) heavily depends on the
quality and the volume of the dataset. However, collecting proper training datasets in many
real-world domains is well-known to be labor-intensive and expensive. Data augmentation
(DA) is a widely accepted solution to improve the low diversity datasets. But selecting optimal DA technique combinations (DA policies) based on the given dataset using traditional trial-and-error approach is time-consuming and requires domain expertise.
In this study, the author attempts to automate the process of selecting optimal DA policies
based on the given dataset by further enhancing Automated Data Augmentation (AutoDA)
technologies. To achieve this goal, the author redefined the traditional AutoDA search space
by reducing the number of hyper-parameters and proposing novel search space exploration
strategy that utilizes neural network and gradient decent technologies. Through these
modifications, the author was able to bridge the existing research gap of achieving efficiency and effectiveness of AutoDA solutions at the same time."
