Improved Data Modeling Using Coupled Artificial Neural Networks

Our senses perceive the world, but what happens if one of the senses is degraded through illness or injury? In such situations, the brain compensates by enhancing the remaining senses. This suggests that networks that process the data received by the senses are coupled. Similar situations can occur in scientific and engineering problems when independent measurement methods, based on different principles, are used to study the same characteristics of a system. In such situation, one can develop reliable artificial neural network (ANN) based models; each trained using data obtained by a different measurement method. This raises the question if it is possible to couple these different models to obtain and improved more accurate model. In this paper, we explore this possibility by training two ANN models that can recognize alphabet letters in a noisy environment. The performance of these ANNs are optimized by varying the number of hidden neurons (HN). The first ANN model trained using pictorial presentation of the letters while the second by corresponding audio signals. The two separate ANNs are trained using the two alphabet letters presentation to which different levels of white noise are added. Different schemes to couple the two systems are examined. For some coupling schemes, the combined system result in highly improved letter recognition than the two original separate ANNs did. Examination of the entropy related to the number of HNs showed that increased entropy is related to a higher error in letter recognition.