Running things through lossy compression is transformative?

The quotation marks indicate that _I_ don't think it is. Especially given that modern deep learning is over-paramaterized to the point that it interpolates training examples.