Why add a dropout layer leading to large val_loss and val_mae?

Question

Recently, I use several dense layers for a regression task.

I use the SELU activations with lecun_normal for each dense layer.

However, when I add one AlphaDro