Analog feedforward neural networks with very low precision weights
S.A. Alibeik, F. Nemati, M. Sharif-Bakhtiar · 2002
An off chip training algorithm for feedforward neural networks is presented. This algorithm has been successfully used to train networks with weight precision as low as 1 bit. The effect of reducing the weight precision on the generalization ability of the network is presented. The network performance, in the presence of hardware non-idealities, has also been investigated. It is shown that a network with low precision weights can well tolerate the effect of hardware non-idealities if the network is properly trained.