Inner workings of tabular NN

Hi Everyone,
I trying to understand the initializations and parameter updates for tabular model using NN. Image below is from the chapter Collab_filter in 2020 fastbook. When we are using the NN I see that we are using standard linear layers from torch. In this model are the weights still initialized for the model (not the embeddings). I am having a hard time understanding the calculation behind the scene.

  1. Is it just the embeddings that are used during calculation like typical collaborative filtering models?
  2. Or do the weights and bias get initialized for the linear layers as well? This would result in two sets of random parameters (embeddings and parameters for the linear layers)
  3. If it is the second option then do both the embeddings and the weights and biases of the linear model get updated?
  • I tried to investigate the actual numbers: the embeddings seem to change before and after training. I was unable to track down the parameters for the linear layers.
  • I appreciate your assistance.
    — If this all sounds confusing: I am just trying to understand the calculation and parameter update of the forward method while using NN in collab_filtering.
    Thank you very much for your help! :pray::smile: