Towards the Design of Bithreshold ANN Regressor

The problem of the application of the multithreshold approach in the machine learning and the neural computation is considered in the paper. We design the model of hybrid 2-layer neural network whose hidden layer consists of bithreshold neural units and employs the winner-take-all computing principle. This model can be useful in the solving regression problems. The model-based approach to the supervised learning of such networks is proposed. Two hyperparameters of the learning algorithm allow to adjust the approximation capability of the network and to tune its ability to generalization. We present the simulation results on the benchmark dataset, which confirm that proposed learning procedure yields the bithreshold neural network that is suitable to make predictions when working with real-world data.