Digital neurochip for fully-connected neural network
N.A. Umnow · 1992
The author presents the design of a digital neurochip architecture. A fully connected neural network with the error backpropagation training method is assumed. The training and the working time of the neural network, when emulated on a conventional computer, shows a power-of-two dependence on the number of neurons. The neurochip in combination with the computer provides a linear dependence.>