Signal processing with delta-sigma modulated bitstreams for artificial neurons

Delta-sigma modulated bitstreams are used to implement the multiply-and-accumulate operation and the rectifying linear unit (ReLU) activation function, essential building blocks in artificial neurons. Our simulation shows that the accuracy is improved by using delta-sigma modulated bitstreams compared to other bitstreams used in the stochastic computing (SC). Specifically, as the bitstream length
$N$
increases, the root-mean-square error (RMSE) improves in proportion to or even better than
$1/N$
, which is due to the noise-shaping characteristics inherent in the delta-sigma modulation. On the other hand, in the SC scheme, the RMSE improvement is only proportional to
$1/\sqrt{N}$
.