Transfer learning capability of Evolving Logic Network

This paper proposes a new kind of transfer learning scheme based on ELN and its learning algorithm. ELN was originally proposed as only a fast incremental learning scheme, which resolves a stability-plasticity dilemma and achieves its fast learning speed by storing and reusing old knowledge in learning of new knowledge. Its knowledge management was originally designed for incrementally learning multiple data sets of a single problem; however, it can also be used in transfer learning for multiple problems. ELN learning algorithm can store the old knowledge in a form of sub-maximally consistent regions and get new knowledge for a new problem by combining this old knowledge, obtained from the previous learned problems. With this reuse of old knowledge, new knowledge can be learned more easily, that is, with a smaller number of learning steps and from a smaller number of data, as more problems become learned and more knowledge becomes accumulated. To validate the ELN learning as a transfer learning, we conducted some experiments and investigated the learning performance improvement as more problems are learned.