Input Layer Binarization with Bit-Plane Encoding

Binary Neural Networks (BNNs) use 1-bit weights and acti- vations to e ciently execute deep convolutional neural networks on edge devices. Nevertheless, the binarization of the rst layer is conventionally excluded, as it leads to a large accuracy loss. The few works address- ing the rst layer binarization, typically increase the number of input channels to enhance data representation; such data expansion raises the amount of operations needed and it is feasible only on systems with enough computational resources. In this work, we present a new method to binarize the rst layer using directly the 8-bit representation of input data; we exploit the standard bit-planes encoding to extract features bit-wise (using depth-wise convolutions); after a re-weighting stage, fea- tures are fused again. The resulting model is fully binarized and our rst layer binarization approach is model independent. The concept is evalu- ated on three classi cation datasets (CIFAR10, SVHN and CIFAR100) for di erent model architectures (VGG and ResNet) and, the proposed technique outperforms state of the art methods both in accuracy and BMACs reduction.