(likely an artifact of most of academia focused on finding models conditioned on standard datasets)
Prikaži ovu nit
We see more significant improvements from training data distribution search (data splits + oversampling factor ratios) than neural architecture search. The latter is so overrated :)
-
-
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Could you explain how to go about this(training data distribution search) or point to a paper that does?
-
There's no paper. That's exactly the problem.
- Još 3 druga odgovora
Novi razgovor -
-
-
What is training data distribution search? (A Google search for that exact string turns up zero results!)
-
Haha yes, because I just made it up. It's the subtle art of composing batches of data during training. (would love to call it science one day instead)
- Još 8 drugih odgovora
Novi razgovor -
-
-
The Efficient Backprop paper (1998!) contains a lot of hidden gems that can give insights in this context https://cs.nyu.edu/~yann/2007f-G22-2565-001/diglib/lecun-98b.pdf …pic.twitter.com/k9OwaB01kQ
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Open Long-Tailed Recognition: https://bair.berkeley.edu/blog/2019/05/13/oltr/ … .
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
You mean creating better batches? How's this? http://arxiv.org/abs/1909.08700 (upcoming in conll). Prime batch sizes for the win.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
this paper compares only batch sizes without taking the labels of datapoints into account : http://jmlr.org/papers/volume20/18-789/18-789.pdf …
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
What do you mean by "oversampling factor ratios"?
-
Presumably oversampling rare classes so that they are weighted more heavily than they would be otherwise. Possibly also looking at the confusion matrix to preferentially study the items you perform worst at?
- Još 1 odgovor
Novi razgovor -
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.