I’m implementing a neural model for a task where part of the input is supposed to be several categorical variables. At inference time, I wish to embed each attribute separately and then take the mean.

If I average the embeddings at train time similarly as during inference, will a model be able to learn each embedding correctly? I.e. will the model be able to figure out which embedding to update and how even if only a mean of them was fed as input? My intuition tells me that MeanBackwards autograd is supposed to help with this but I’m not sure I understand exactly how that is going to work!