Performance of finite iteration DTCNN with truncated stationary templates [digit recognition example]

In this paper, we consider finite-wordlength effects in a finite-iteration DTCNN (discrete time cellular neural network). Using the digit recognition example, we demonstrate that it is possible to effectively design templates with limited precision. Further, using the digit classification example, we show that the performance is not affected either by truncation to two decimal places in the learning phase/design of templates or the finite precision (8-bit) implementations of the templates.