POLYCiNN: Multiclass Binary Inference Engine using Convolutional Decision Forests

Abstract

Convolutional Neural Networks (CNNs) have achieved significant success in image classification. One of the main reasons that CNNs achieve state-of-the-art accuracy is using many multi-scale learnable windowed feature detectors called kernels. Fetching of kernel feature weights from memory and performing the associated multiply and accumulate computations consume massive amount of energy. This hinders the widespread usage of CNNs, especially in embedded devices. In comparison with CNNs, decision forests are computationally efficient since they are composed of decision trees, which are binary classifiers by nature and can be implemented using AND-OR gates instead of costly multiply and accumulate units. In this paper, we investigate the migration of CNNs to decision forests as one of the promising approaches for reducing both execution time and power consumption while achieving acceptable accuracy. We introduce POLYCiNN, an architecture composed of a stack of decision forests. Each decision forest classifies one of the overlapped sub-images of the original image. Then, all decision forest classifications are fused together to classify the input image. In POLYCiNN, each decision tree is implemented in a single 6-input Look-Up Table and requires no memory access. Therefore, POLYCiNN can be efficiently mapped to simple and densely parallel hardware designs. We validate the performance of POLYCiNN on the benchmark image classification tasks of the MNIST, CIFAR-10 and SVHN datasets.