Accelerating Training of Deep Neural Networks via Sparse Edge Processing