Comment by m_ke

We really need new hardware optimized for sparse compute. Deep Learning models would work way better with much higher dimensional sparse vectors but current hardware only excels at dense GMMs and structured sparsity.