A new approach for continous Learning

Abstract

This paper presents a new method for continous learning based on data transformation. The proposed approach is applicable where individual training datasets are separated and not sharable. This approach includes a long short term memory network combined with a pooling process. The data must be transformed to a new feature space such that it cannot be converted back to the originals, while it can still keep the same prediction performance. In this method, it is assumed that label data is sharable. The method is evaluated based on real data on permeability prediction. The experimental results show that this approach is sufficient for continous learning that is useful for combining the knowledge from different data sources.