CVPR Daily - Tuesday

He explains: “ Accuracy with deep learning models is important, and data augmentation is an easy but impactful way of improving the accuracy . A lot of the work is focused on architectural improvements, rather than processing data. Since experts already have an idea of what symmetries are, and it’s clear how they can utilise those symmetries to augment the data, it would be good to increase the accuracy. Something else we have been seeing recently that we didn’t necessarily think about when we were writing the paper is that it seems that increased data augmentation helps a lot with robustness. Recently there was an ICML workshop here on robustness for machine learning models to common corruptions and noise. What several people found is that AutoAugment leads to the best robustness . Although it was actually only trained for increased validation accuracy. ” In terms of next steps, Dogus says that they want to apply this idea to other domains. Not just image classification, but video and object detection , for example. They recently had a paper, SpecAugment , about applying AutoAugment to speech to benefit speech rd to get to work. Since their work, other people have reproduced it with Bayesian optimization or with population-based training or a few other methods, and Dogus is encouraged by how easy it is to get the same results with different optimizers. 11 DAILY CVPR Tuesday AutoAugment To learn more about this work, you are encouraged to come along to Dogus’s oral [1.1A] today at 10:04 and poster [12] at 10:15-13:00. “ Accuracy with deep learning models is important, and data augmentation is an easy but impactful way of improving the accuracy!”