Which way is up?

Maybe a not completely serious question but nonetheless I still feel this deserves clarification once and for all :slight_smile: (also, maybe more people are similarly confused).

I always envisioned neural networks as a stack of layers, not unlike a pyramid. Logically, input would be on the bottom and output on top. But reading fastai code we seem to be referring to the top_model as being the part of the model that takes in inputs. I think I also came across similar naming convention in one of the papers.

Could I please ask if this is generally the naming convention to call the end of the model that takes in inputs the top and the predictions come out the bottom?


Heh - I always think of them the same way you do. Not sure why I used that attribute name…


May be because when you start implementing a neural net in code, you start with the input layer followed (downwards) by other layers. That makes the end of the model that takes in inputs to be on top.