Edge artificial intelligence (AI) is receiving a tremendous amount of interest from the machine learning community due to the ever-increasing popularization of the Internet of Things (IoT). Unfortunately, the incorporation of AI characteristics to edge computing devices presents the drawbacks of being power and area hungry for typical deep learning techniques such as convolutional neural networks (CNNs). In this work, we propose a power-and-area efficient architecture based on the exploitation of the correlation phenomenon in stochastic computing (SC) systems. The proposed architecture solves the challenges that a CNN implementation with SC (SC-CNN) may present, such as the high resources used in binary-to-stochastic conversion, the inaccuracy produced by undesired correlation between signals, and the complexity of the stochastic maximum function implementation. To prove that our architecture meets the requirements of edge intelligence realization, we embed a fully parallel CNN in a single field-programmable gate array (FPGA) chip. The results obtained showed a better performance than traditional binary logic and other SC implementations. In addition, we performed a full VLSI synthesis of the proposed design, showing that it presents better overall characteristics than other recently published VLSI architectures.
30 May 2014 in International Journal of Neural Systems
Nowadays Internet of Things (IoT) and mobile systems use more and more Machine Learning based solutions, which implies a high computation cost with a low energy consumption. This is causing a revival of interest in unconventional hardware computing methods capable of implementing both linear and nonlinear functions with less hardware overhead than conventional fixed point and floating point alternatives. Particularly, this work proposes a novel Radial Basis Function Neural Network (RBF-NN) hardware implementation based on Stochastic Computing (SC), which applies probabilistic laws over conventional digital gates. Several complex functions design to implement RBF-NN are presented and theoretically analyzed, such as the squared Euclidean distance and the stochastic Gaussian kernel similarity function between input samples and prototypes. The efficiency and performance of the methodology is tested over well-known pattern recognition tasks, including the MNIST dataset. The results show a low-cost methodology in terms of logic resources and power, along with an inherent capability to implement complex functions in a simple way. This methodology enables the implementation of massively parallel large scale RBF-NN with relatively low hardware requirements while maintaining 96.20% accuracy, which is nearly the same for the floating point and fixed point models (96.4% and 96.25%, respectively).