Trained neural networks have successfully been applied for channel equalization and decoding. The structure of some dynamical neural networks is especially suitable for certain tasks, like using Hopfield networks as multiuser detectors. In this case no training is required because of the identical structure of the network and the task. High-order neural networks have been successfully applied in pattern classification, usually producing better performance, both in approximation accuracy and computational cost, than conventional neural architectures such as the multilayer perceptron (MLP). Hopfield has shown that a system of highly interconnected “neurons” has useful collective computational properties. The mathematical modelling has been based on “neurons” that are different from a realistic functioning of a simple electronic circuit. Hopfield proved that a modified model built of operational amplifiers and resistors, has the same collective properties like the original model. The new model “the analog model” has a quadratic energy function and has been applied to many optimization problems like the travelling salesman problem and the analog to digital conversion, among others. Researchers aim to avoide the limitation of this model by increasing the order of the energy function and thus allowing nonlinear combination of the outputs as feedback inputs of the neuron. In this work a continuous modelling of iterative threshold decoding for binary linear codes is presented. We want to investigate whether the iterative threshold decoding algorithm matches well with the structure of a continuous high-order recurrent neural network. The performance of the continuous modelling will be evaluated by simulations and will compared with the corresponding digital modelling. The motivation of this work is that continuous modelling conjuncted with the high developement in the realization of analog circuits might improves the power/speed ratio.