Black Box Image Augmentation:
In this project, I explore how to augment data with synthetically produced images using a black-box algorithm. Through a feedback loop and with Bayesian Optimization, I allow the model to self-identify critical image features for it to be able to classify images. As a result, I found that the black box augmentation transforms the original image to a hyper-vibrant image and boosts the model’s accuracy with very few datapoints needing to be added to the training set.
My work on leveraging Bayesian optimization for synthetic data augmentation has been accepted and invited to many industry conferences including MLConf and Nvidia’s GTC.
Leveraging Bayesian Optimization for Synthetic Data Augmentation on low resource image datasets
Examples of auto-augmented images
Watch the talk
Resources
-
To recreate or repurpose this work please use this repo.
-
To play around with the SigOpt dashboard and analyze results for yourself, take a look at the experiment
-