Considering that random forests have many layers and beat deep learning in most applications, maybe we just need to rebrand them as deep forests and they'll be the next big thing.
-
Sure, I mean we should all be surprised that decision trees didn't solve the difficult tasks and weren't behind the likes of alpha fold, multilingual machine translation, image-to-text generation, self-driving cars etc.
Deep learning is generally better for problems with known a priori structure (vision, language) and worse for those without. Check Kaggle, for example. Everyone should know that.
-
-
What is a known a priory structure?
-
I think it's that you know the general shape of data before doing anything (tensors for images; vectors of text) and adapt layers and parameters to fit those structures. Hit the training button and watch machine go brrrr.
Keskustelun loppu
Uusi keskustelu -
-
-
Random Forests & GradBoost dominate tasks on Kaggle but IMO the strides DNNs have made on representing fundam. tasks in vision, speech, lang its not surprising why they've gained traction. I agree built in structure (convolutions & attention) have driven much of this success.
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
-
-
Most tabular data based problems were predominantly handled by GBDTLR like approaches but shifting that to leverage DL has been very positive, especially for our Business Unit. I'm sure others in recsys would be able to +1 this.
Kiitos. Käytämme tätä aikajanasi parantamiseen. KumoaKumoa
-
Lataaminen näyttää kestävän hetken.
Twitter saattaa olla ruuhkautunut tai ongelma on muuten hetkellinen. Yritä uudelleen tai käy Twitterin tilasivulla saadaksesi lisätietoja.