Neural training and generalisation of sequences using continuous temporal structure

An approach for sequential neural behavior called continuous backpropagation is evolved from standard backpropagation where a state is replaced by a state transition sequence as the goal weight condition. The approach may be used to train mappings of analog input/output (I/O) signals or discrete I/O sequences with underlying continuity. An arbitrarily increasing number of values in the I/O sequences may be trained without having to increase the number of hidden units. The training and generalization techniques are illustrated by a sequential four-spiral version of Wieland´s two-spirals problem. The results show a substantial improvement over standard state-based backpropagation