An Improved Incremental Classifier and Representation Learning Method for Elderly Escort Robots

Elderly escort robots are gradually becoming one of the important roles of elderly care services. The problem that restricts the wide application of elderly escort robots in practical scenarios is how to make them have the ability of continual learning, so that they can autonomously learn and respond in dynamic service scenarios. To this end, we propose a continual learning network based on an improved Incremental Classifier and Representation Learning (iCaRL) method for robots. The network uses replay and regularization strategy to train a ResNet with incremental classes. We mainly focus on improve the replay method in two aspects. Firstly, a density-peaks-based prototype selection strategy is proposed, which allow the network to obtain accurate prototype for arbitrary-shaped data distribution. Secondly, A self-organizing incremental neural network is introduced for each class as exemplar memory to replay in the ResNet training process. It can not only learn generalized representations of each class to enhance the diversity of its exemplars, but also learn any number of classes due to its dynamic architecture. Experimental results demonstrate that our method can achieve better learning effectiveness and efficiency over several state-of-art algorithms.