Papers of the day   All papers

AutoAugment: Learning Augmentation Strategies from Data

Comments

Google AI: Introducing SimCLR, a new approach to self- and semi-supervised learning that can learn good image representations without requiring human annotations, and achieves a new state of the art on ImageNet when fine-tuned on only 1% labeled data. Learn more at https://goo.gle/3e2owMa https://t.co/6gGvLEkpxp

12 replies, 1020 likes


Xander Steenbrugge: Self-supervised learning is starting to work **very** well thanks to contrastive losses: @ylecun's cake🥮 is growing! Next step: combine this with fully learnable augmentations (~https://arxiv.org/abs/1805.09501) + insight into how these influence the final representations wrt new tasks

2 replies, 153 likes


hardmaru: Papers excluding first (NAS) and last point: • better activation functions https://arxiv.org/abs/1710.05941 • better learning rules https://arxiv.org/abs/1709.07417 • better data augmentation https://arxiv.org/abs/1805.09501 • better loss functions https://openai.com/blog/evolved-policy-gradients/ & https://arxiv.org/abs/1905.11528

3 replies, 152 likes


William Hsu: Interesting use of reinforcement learning to optimize data augmentation. Instead of performing a fixed number of transformations, a search is used to identify a set of operations and their ordering that yields the best validation accuracy. #CVPR2019 https://arxiv.org/abs/1805.09501

0 replies, 1 likes


Content

Found on Apr 08 2020 at https://arxiv.org/pdf/1805.09501.pdf

PDF content of a computer science paper: AutoAugment: Learning Augmentation Strategies from Data