Data Tech

Training dataset sizes are a critical factor in the creation of accurate neural network image classifiers where the axiom “bigger-is-better” appears to hold true. Image augmentation (creating several instances of an original image by applying image transformations) is often used to increase the size of datasets when the original dataset is too small. Although several augmentation strategies have been developed that have successfully improved neural network performance, little research has been done to study the appropriate ratio of augmented to original data. In this presentation, Tom will introduce a search strategy that uses statistical mixture experiments to identify the optimal blend of several different image augmentation methods. He will also discuss how hyperparameter tuning can be incorporated into this process to simultaneously tune hyperparameters and augmentation strategies for efficient deep learning model optimization. The presentation concludes with a case study where a mixture experiment was used to identify the optimal augmentation strategy for a neural network used for manufacturing visual defect detection, resulting in a significant improvement in performance on a validation dataset.