Question regarding using deep learning for tabular model

I have a question regarding tabular model.

  1. What are the parameters updated every batch?
  2. Are the random initializations of the embeddings and the weights and biases (also random) of the linear layers updated?
  3. Is the computation in forward method, matrix multiply between the weights of the layers and the imbedding matrices?
  4. If no for 3, can you please explain the computation happening at method forward and the parametrs updated following that.
    Thank you! :pray::smile: