Temporal Sequence Pattern Learning and Dynamic System Control (DSC)

We have investigated the role of temporal sequence learning, using an unsuper- vised artificial neural network (1), called Monoconnected Autoreflexive Neural Net- work, for better understanding the implicit learning process role, involved during elementary associative learning processes. Several neural network models have been proposed to describe implicit learning (IL), using unsupervised and self-organized models (2, 3). In our experiments we used a real biochemical data set consisting of 15 features, that deals with penicillin production (112 temporal sequence blocks with 11 sequence points per block (1232 patterns). The prediction task requires that the neural network can predict the correct sequence position, after a preliminary training (50% of all patterns). After training, the neural network learn to find the correct position in the temporal sequences with good accuracy. Our results seem to confirm that elementary associative learning, could be used in temporal sequence learning and that dynamic system control (DSC) tasks (for instance to know which features are more sensitive to a better penicillin production), could be derived from the implicit learning process, using the importance of different features, recovered from the weight matrix analysis.