LSF restoration by means of a neural network

The LSF restoration problem is written as a Maximum Entropy one, where the constraint on the restoration energy is dictated by the “Discrepancy Principle”. The ME solution is found by means of a continuous-Hopfield neural network which reduces the energy of the output misfit, and maximizes the restoration entropy at the same time. A positive learning parameter controls the constraint compliance. Prior knowledge insertion into the netʹs algorithm, such as prior LSF models, upper bounds, etc. is presented. Simulations, both with computer generated and experimental data are carried out. The results are compared to those of the Least Squares method. Sensitivity of constraint fulfillment is analyzed.