Fast arithmetic computing with neural networks

Abstract

The authors introduce a restricted model of a neuron which is more practical as a model of computation then the classical model of a neuron. The authors define a model of neural networks as a feedforward network of such neurons. Whereas any logic circuit of polynomial size (in n) that computes the product of two n-bit numbers requires unbounded delay, such computations can be done in a neural network with constant delay. The authors improve some known results by showing that the product of two n-bit numbers and sorting of n n-bit numbers can both be computed by a polynomial size neural network using only four unit delays, independent of n . Moreover, the weights of each threshold element in the neural networks require only O(log n)-bit (instead of n-bit) accuracy.