Merging information in the data and weight spaces

The paper addresses the problem of combining independent information which can be available in both the data and parameters spaces: the objective is to obtain a neural model which takes into account the information available from both sources. The problem is approached in the framework of the probabilistic interpretation of neural modelling and the indetermination associated to the training process is taken into account by considering an appropriate distribution in the weight space associated to each solution vector. A computationally light procedure is proposed to merge the information associated to the different solutions. The effectiveness of the proposed procedure is shown by means of experiments of feedforward neural networks for classification tasks