Accelerating Deep Neural Networks on Low Power Heterogeneous Architectures