News for hackers. Stuff that matters.
Re:The Transformer Family by jerpint on Monday January 30, 2023 @06:50AM (#34578536)
There are bit caveats here:

Vision transformers outperform CNNs in HUGE data regimes. On small datasets, CNNs still shine.

Also, if you take a CNN with modern tricks, they can be on par with vision transformers, e.g convnext

Transformers really dominate when you scale the amount of data to infinity

Re:The Transformer Family by alexmolas on Tuesday January 31, 2023 @08:44AM (#34594893)
Yes, that's what I was saying: scale > inductive bias. Apparently, scaling is all you need ;)