arun@vlsi.waterloo.edu (Arun Achyuthan) (02/12/91)
B. Furman and A. A. Abidi of Integrated Circuits and Systems Laboratory of University of California, Los Angles report in the paper titled "An Analog CMOS Backward Error-Propagation LSI" (Proceedings of the 22nd Asilomar Conf. on Signals, Systems and Computers, pp645-648) a fully analog architecture to implement the Back Prop algorithm, which includes on-chip learning. All my simulations to date indicate that the weight change computed in the backward path needs at least 16 bits of resolution and highly linear processing of data, inorder for the algorithm to converge. My knowledge is that both these requirements are very diffcult to be met by analog circuits, due to their inherent properties. Could the authors of the above mentioned paper or anybody else in this newsgroup provide an explanation to this conflicting observation? Arun