Selective visual memory replay with self-evaluation in cognitive robots based on global workspace framework

Learning capability for artificial systems is a well-studied topic, with various schemes enabling the system to develop knowledge continually. Methods based on memory replay are commonly adopted in the literature. This work presents a consciousness-based model integrated with a continual learning scheme for class-incremental learning in visual recognition. We suggest a reciprocal relation between memory maintenance and the learning activity of the system based on psychological evidence. The memory capability fits the continual learning problem. In return, a self-evaluation of knowledge mechanism is proposed for the robot to discriminate the important learning data during interactions to alleviate the memory constraint without degrading the distribution representation of abnormal data. The implemented robotic agent autonomously puts more effort into learning novel knowledge without human intervention. The cognitive architecture based on the Global Workspace Theory for the robotic agent is presented, with which the agent can automatically associate information from different modalities. Memory consolidation is implemented to run in parallel to the memory formation process. The work is validated in a class-incremental object recognition experiment on a robotic agent. The results show that the agent automatically balances the memory distribution for learning and maintains a relatively small set of samples during learning.