I am trying to use the GTNs taught this week for one of my projects. Now, in GTNs, as we know, we have the teacher and the student, wherein, I have the model architecture for the student. Now, I wanted to know, based on the student’s architecture, how can I find suitable architecture for the teacher? Are there any analogies between the student’s and the teacher’s architecture, like they should have the same number of layers, same number of hyperparameters, etc? Also, for a given student’s and teacher’s architecture, how can I decide the number of the outer loop and inner loop iterations?
I hope you all are doing well.
I would like to request you, to take some time and answer this query.
Thank you very much in advance!
This is a nice question. Usually, there are no fixed rules between the architectures. So there are no direct analogies.
However, finding architectures that best suit your data is called Neural Architecture Search (NAS) which you might want to look into. Regarding the same, this paper might be of interest to you:[1912.07768] Generative Teaching Networks: Accelerating Neural Architecture Search by Learning to Generate Synthetic Training Data.
Hope this helps!
I was actually trying to use it in a Research Paper, but I left it, since I was unable to adjust it to my needs. Still, thanks a lot for your response.