Incremental recurrent kernel machines for continuous gesture recognition

Abstract

With the advancement of artificial intelligence technologies in recent years, research on intelligent robots has progressed. Communication robots that engage psychologically with people, in particular, are gaining traction. Hand gesture recognition is important in human-robot interaction because it allows us to comprehend diverse human gestures and their intention. Traditional deep learning approaches need to keep all prior class samples in the system and require model training from the start by combining prior and new examples, which consumes massive amounts of memory and dramatically increases computation cost. In this study, we proposed a method called Incremental Recurrent Kernel Machines (IRKM) that mimics the human lifelong learning process that can continuously learn new gestures without forgetting previously learned gestures. The proposed method consists of two hierarchical memory layers: i) Episodic Memory and ii) Semantic Memory layer. The Episodic Memory layer incrementally clusters incoming sensory data as nodes and learns fine-grained spatiotemporal relationships of them. The Semantic Memory layer adjusts the level of architectural flexibility and generates a topological semantic map with more compact episodic representations based on task-relevant inputs. The generated topological semantic map reflects the memory of the robot in which it is utilized for gesture recognition.