Towards Energy-Efficient Convolutional Neural Network Inference