Is there any work done on studying the biases that maybe introduced by generating a student model by using the top scoring K-examples?
Billion-scale demi-weakly-supervised learning yields record results on image and video recognition. And it's open source. https://ai.facebook.com/blog/billion-scale-semi-supervised-learning …https://ai.facebook.com/blog/billion-scale-semi-supervised-learning …
-
-
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Ah, the public parameter tuning continues! Powered by coal or..?
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
At a high level, semi-weakly-supervised learning relates to our "weakly supervised co-training" framework for text classification, which also uses a teacher-student paradigm. Our
#emnlp2019 paper: https://arxiv.org/pdf/1909.00415.pdf … Excited to apply these ideas for more#NLProc tasks!Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Thanks again for so many fundamental open source contributions. Any general oberservations about when to use weak supervision vs unsupervised learning? Bon weekend!
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
81.2% with ResNet-50 (~4.1B MulAdds) is nowhere near the “state of the art on academic benchmarks for lightweight image and video classification models.” EfficientNet-B4 (4.2B MulAdds) gets 82.6% without any auxiliary training data. https://twitter.com/ylecun/status/1185371540260642822 …pic.twitter.com/t9CwRrOh73
-
I wonder what accuracy could be achieved, if EfficientNet was used as the Student.
Kraj razgovora
Novi razgovor -
-
-
How about clustering visually similar hashtags before feeding to the Teacher model?
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Straight into Chinese surveillance system.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.