Design methodology for neural network simulation of sequential circuits using neural storage elements

Multilayer feedforward networks have been found suitable for applications in which they need to learn binary-to-binary mappings. We propose a design methodology to simulate sequential functions using neural networks. The combinational function is implemented by a perceptron network with single hidden layer trained using an ETL algorithm. Design of neural storage elements similar to flip-flops is also proposed, which are used as memory elements to store the internal states. Use of the ETL algorithm guarantees convergence for any binary-to-binary mapping, and generally leads to faster convergence than the backpropagation algorithm. The resulting network only consists of neural elements, with all the neurons having integer valued weights and activation thresholds making the network more suitable for hardware implementation using digital VLSI technology.