Backpropagation through Time Algorithm for Training Recurrent Neural Networks using Variable Length Instances

Abstract

Artificial Neural Networks (ANNs) are grouped within connectionist techniques of Artificial Intelligence. In particular, Recurrent Neural Networks are a type of ANN which is widely used in signal reproduction tasks and sequence analysis, where causal relationships in time and space take place. On the other hand, in many problems of science and engineering, signals or sequences under analysis do not always have the same length, making it difficult to select a computational technique for information processing. This article presents a flexible implementation of Recurrent Neural Networks which allows designing the desired topology based on specific application problems. Furthermore, the proposed model is capable of learning to use knowledge bases with instances of variable length in an efficient manner. The performance of the suggested implementation is evaluated through a study case of bioinformatics sequence classification. We also mention its application in obtaining artificial earthquakes from seismic scenarios similar to Cuba.