Constructivism-Inspired Meta-Knowledge Learning and Reuse

Efficient continual learning of humans is enabled by interactions of a series of learning mechanisms and human memory system. Constructivism is a theory in education that proposes learners construct knowledge from their experiences, assimilation and accommodation are key to this process. Also, research shows knowledge in human memory system is represented as a high-dimensional and sparse vector. Motivated by these, in our paper, we propose a continual feature learning method based on constructivisim and memory system. We divide continual feature learning into two stages. In assimilation stage, assimilation model captures knowledge that cannot be represented by meta-knowledge in knowledge base and update the knowledge base according to new knowledge. Furthermore, in accommodation stage, accommodation model updates its learning ability based on new tasks and new knowledge base, ensuring the continual feature learning of new tasks without forgetting previous knowledge. Experimental results show our proposed method can further alleviate the forgetting effect of the model.