This is accomplished through 
two primary modifications: a hierarchy of Transformers containing a new convolutional token embedding, and a convolutional Transformer 
block leveraging a convolutional projection.