Does Fastai.Tabular use a deep neural network or a CNN and what is the difference. (sorry if it’s a noob question :p) Any help would be greatly appreciated.
Reading the links links below would help you understand better than I could explain it.
The 3rd link covers some members of the forum, that I really respect @joedockrill and @muellerzr discussing the pros and cons of using CNN’s on tabular data.
As @mrfabulous1 mentions, fastai’s (and most) tabular neural networks are in fact what we would call fully connected networks. This means from start to end every single layer is directly connected. Whereas a CNN (or convolutional neural networks) are not fully connected, with only a few fully connected layers towards the end.
In regards to @mrfabulous1 about recent research, some research has been done with tabular data and CNN’s. When I looked at it ~1 year ago I was not impressed as they were mostly just writing out the various data into a picture and it worked (for the most part). The results weren’t outstanding. There’s been some recent developments though that look promising (mostly using T-SNE or various other methods). A few papers and kernels to peek at are: