Cross-Domain Knowledge Transfer Using Structured Representations

Proceedings of the AAAI Conference on Artificial Intelligence, 21

Volume

Technical Papers

Machine Learning

Previous work in knowledge transfer in machine learning has been restricted to tasks in a single domain. However, evidence from psychology and neuroscience suggests that humans are capable of transferring knowledge across domains. We present here a novel learning method, based on neuroevolution, for transferring knowledge across domains. We use many-layered, sparsely-connected neural networks in order to learn a structural representation of tasks. Then we mine frequent sub-graphs in order to discover sub-networks that are useful for multiple tasks. These sub-networks are then used as primitives for speeding up the learning of subsequent related tasks, which may be in different domains.