Memory-Optimal Direct Convolutions for Maximizing Classification Accuracy in Embedded Applications