Position: A Theory of Deep Learning Must Include Compositional Sparsity | Read Paper on Bytez