Submitted to The Third Swedish Symposium on Deep Learning