Stochastic logics with two-dimensional state transfer structure and its application in the artificial neural network

Stochastic computing is a special algorithm that performs mathematical operations with probabilistic values of bit streams rather than traditional deterministic values. The main advantage of stochastic computing is its great simplicity of hardware arithmetic units for mathematical operations to reduce the circuit cost. This paper discusses the principle of the stochastic computing and its main arithmetic logic. It analyzes a two-dimension state transition topology structure, and discusses the Gaussian function implementation method based on the two-dimension Finite State Machin (FSM). Then, a low cost stochastic radial basis function neural network model is proposed. Results from two pattern recognition tests show that the difference of the mean squared error between the stochastic network output value and the corresponding deterministic network output value can be less than 1.3%. FPGA implementation results show that the hardware resource requirement of the proposed stochastic hidden neuron is only 1.2% of the corresponding deterministic hidden neuron with the interpolated look-up table, and is 2.0% of the CORDIC algorithm. The accuracy, speed and power of the stochastic network can be tradeoff dynamically. This network is suitable for the low cost and low power applications like embedded, portable and wearable devices.