Achieving Goal-directed Cognitive Tasks by Coordinating Visual Attention, Recognition and Action（セミナー・シンポジウム）

To achieve visually-guided actions for multiple object manipulation requires proactive sequential visual attention shifts and visual recognition synchronized with adequate accompanying hand movements. In this case, the selective visual attention and recognition model continuously catches the visual environment, which contains multiple objects, in order to perceive the current relationship between a human and the environment. By sequentially perceiving the characteristics and localization of target objects, human beings can easily generate a suitable behavior according to a given task. Thus, behavior causes changes in the environment, which in turn lead to different visual perception results, and this cycle continues until the goal-directed cognitive tasks are achieved. Based on the understanding of those cognitive functions, I will present artificial cognitive functions to develop an autonomous robot system with human-like characteristics. First, a selective visual attention model is presented that uses bottom-up visual features and previously acquired top-down knowledge based on understanding the visual what and where pathway in the human brain in order to focus on a specific salient object or area. Second, an object recognition model is presented based on incremental feature representation and a hierarchical feature classifier that offers plasticity to accommodate additional input data and reduces the problem of forgetting previously learned information. Based on these two visual specific cognitive functions, goal-directed behavior generation in environments involving multiple objects is studied using the action-perception cycle with dynamic neural networks. The model is evaluated by neuro-robotic experiments that include behavior tasks involving multiple objects.

Sungmoon Jeong got the Ph. D. from Kyungpook National University, Korea in 2013, and is currently an Assistant Professor at School of Information Science, Japan Advanced Institute of Science and Technology, Nomi, Japan. His research interests include cognitive system, intelligent signal processing, incremental learning, pattern recognition, and real time application systems.