Meta-learning for Fast Incremental Learning

Model based learning systems usually face to a problem of forgetting as a result of the incremental learning of new instances. Normally, the systems have to re-learn past instances to avoid this problem. However, the re-learning process wastes substantial learning time. To reduce learning time, we propose a novel incremental learning system, which consists of two neural networks: a main-learning module and a meta-learning module. The main-learning module approximates a continuous function between input and desired output value, while the meta-learning module predicts an appropriate change in parameters of the main-learning module for incremental learning. The meta-learning module acquires the learning strategy for modifying current parameters not only to adjust the main-learning module’s behavior for new instances but also to avoid forgetting past learned skills.