Stochastic connection neural networks

We investigate a novel neural network model which uses stochastic weights. It is shown that the functionality of the network is comparable to that of a general stochastic neural network using standard sigmoid activation functions. For the multilayer feedforward structure we demonstrate the network can be successfully used to solve a real problem like handwritten digit recognition. It is also shown that the recurrent network is as powerful as a Boltzmann machine. A new technique to implement simulated annealing is presented. Simulation results on the graph bisection problem demonstrate the model is efficient for global optimization.