Design Of Approximate Addition Unit For Binarized Weight Network

Deep Neural Network(DNN),as an excellent algorithm to replace traditional algorithms,has shown significant improvements.In some cases,their accuracy even exceeds humans.However,the price of high accuracy is higher computational complexity.In order to make the DNN more lightweight,quantifying the activation value and weight value has become the common methods.Among them,the Binarized Weight Network(BWN)has only addition operations because of reducing the weight to 1-bit.Since the activation value is still 16-bit,the DNN can be further optimized,which makes it possible to apply the approximate calculation in the BWN.Firstly,the existing approximate adders are compared and analyzed,then the single-bit structure of approximate full adders are analyzed.Monte Carlo simulation results show that the approximate full adders are superior to other types of approximate adders in terms of Power Delay Product(PDP)and Mean Relative Error Distance(MRED).In order to make the evaluation more efficient,an approximate full adder error evaluation method based on the probability model is proposed,which can quickly evaluate the error characteristics of approximate adders.Then,the approximate addition unit is extended to the approximate accumulation of multiple numbers by the approximate addition unit error model based on approximate noise.Based on this theory,an approximate calculation architecture for the BWN is designed.In addition,the precision control method is proposed to control the precision dynamically when the circuit is running.Finally,the BWN approximate calculation unit is applied to the keywords spotting system based on the BWN.Through a series of analyses in the original network data,the approximate calculation circuit structure of the BWN is determined.Under the process of TSMC’s 22 nm ULL UHVT,0.72 V voltage and 25°C,compared with the keywords spotting system without the approximate calculation unit,the overall power consumption of the keywords spotting system has decreased by 11.5%,while the network recognition accuracy has only decreased by 0.3%.