Prof. Anima Anandkumar on Twitter: "2/Instead, we use proxy guidance for synthetic training to avoid catastrophic forgetting. We use learning to optimize to tune layer-wise learning rates. This allows us to get SOTA on VisDA (classification) and Cityscapes (segmentation) tasks. Also improves model attention."

Country	Code	For customers of
United States	40404	(any)
Canada	21212	(any)
United Kingdom	86444	Vodafone, Orange, 3, O2
Brazil	40404	Nextel, TIM
Haiti	40404	Digicel, Voila
Ireland	51210	Vodafone, O2
India	53000	Bharti Airtel, Videocon, Reliance
Indonesia	89887	AXIS, 3, Telkomsel, Indosat, XL Axiata
Italy	4880804	Wind
3424486444	Vodafone
» See SMS short codes for other countries

1/Our #ICML2020 paper today: Automated Syn-to-Real Generalization. @ZhidingYu @chenwydj @NVIDIAAI Previous works: start with pre-trained Imagenet model, train with synthetic data, do early stopping / learning rate tuning to previous catastrophic forgetting. https://twitter.com/ZhidingYu/status/1283384464576372737 …

This Tweet is unavailable.

1 reply 6 retweets 43 likes

Show this thread

Tweets