Is it
- number of units in a layer
- number of layers
- overall complexity of the network (both 1 and 2)
$\begingroup$
$\endgroup$
1
-
$\begingroup$ Usually it refers to the number of layers that are sequential (i.e. one after the other). However, it isn't used much like a quantity (e.g. this network has a depth of 8), rather a qualitative characteristic of the network (e.g. this network is deeper than the other). $\endgroup$ – Djib2011 Feb 18 '20 at 9:33
Add a comment
|
$\begingroup$
$\endgroup$
Very good question, I have heard it refered to all the 3 ways you described. I am not sure if there is a purely objective answer so I will answer with what I understand by "depth" generally:
- If speaking of model architecture: the number of layers of the model.
- If speaking of CNN layers: the number of filters (not to be confused with the size of the kernel).
- If speaking of RNN layers: the number of time steps (temporal "depth").
On the other hand I have never heard about depth regarding to FC layers. As I said the term "depth" is very subjective depending on the speakers. It should not be this way, and maybe theoretically is not, but I speak from what I heard in my experience.
