Adaptive Precision Cellular Nonlinear Network

The versatility of cellular nonlinear network (CNN) made it attractive in various scientific and engineering applications. The strength of solving problems using CNN comes from the fact that it deals with nonlinear dynamic systems. In this paper, feasible criteria for adaptive hardware approximation, automated method to switch between high and low precisions, is proposed to efficiently reduce energy consumption of a digital CNN. Two different criteria, state- and sign-dependent control, are presented and compared by simulating discrete-time CNN with various applications: image processing, associative memory, and scientific computing. Simulation results show that the state-dependent control is better suited for automatically introducing approximate computing in CNN accelerator. Even further, this control method allows more aggressive power reduction by stopping the computations for those cells in the saturation region of CNN dynamics for applications like associative memory. As a result, adaptive hardware approximation is effective in reducing energy consumption of a digital CNN compared with simply injecting predetermined approximation for entire CNN computations. The experimental results also show that the acceptable level of energy reduction differs from the application, equivalently the type of CNN.