Skip to Main Content
Feedforward and recurrent neural networks have enjoyed great popularity in the field of machine learning. Hardware implementation of these networks, often using a backpropagation algorithm for learning, has many advantages over software implementation, the main advantage being speedup due to taking advantage of the inherent parallelism of neural networks. However, tradeoffs are often introduced with respect to speed, area, precision, and other factors. Moreover, many of the calculations involve multiplications, which can be costly and complex to implement in custom hardware. This paper proposes a piecewise linear approximation architecture that can be used to replace multiplications with a series of shifts and additions. This, combined with the same approximation for the nonlinear activation function and its derivative, results in a neural network realization in which all arithmetic and function computations are carried out using adders only.