The 1990 paper “A Stochastic Version of the Delta Rule” by Stephen J. Hanson introduces a modified version of the delta rule incorporating synaptic noise into the weight update process of neural networks. This stochastic approach, termed the Stochastic Delta Rule (SDR), treats each weight as a random variable characterized by a mean and standard deviation. By adding controlled noise, SDR enhances the network's ability to escape local minima, thereby improving convergence speed and reliability compared to standard backpropagation methods of the time.