Synergic Attention-based Visuospatial Episodic Memory and its Application to Wearable Navigation

在这项研究中，我们提出了基本的方法，为伙伴机器，如熟悉的计算机和机器人，支持人类所有者的日常活动，通过关注他们的日常空间，并与他们分享注意力和关注。这些方法包括表示和控制的主人和他/她的合作伙伴机器之间的协同注意力，学习的注意力结构的竞争神经网络的基础上，和基于贝叶斯网络的个人建模和现实世界的信息导航使用视觉注意和响应话语的主人。在协同注意方面，我们建立了一种通过主人关注诱导的主动空间注意来调节显著性驱动注意的方法。在此基础上，建立了一种注意力转移与控制方法，提出了一种主机器与伙伴机器之间的协同注意力控制方法。在注意结构的学习方面，我们建立了一个将某一场景中的一组注意点编码为注意的模型 ...更多信息 结构代码该模型由竞争神经网络COGNET组成，COGNET对关注点中包含的对象进行编码，并对关注点的大小和位置进行编码。通过对COGNET的快速自组织学习和对关注点内物体的扫视识别等主要特征的实验评价，证实了注意结构编码对于视觉空间事件的编码是足够有用的。至于个人建模和基于它的信息提供，我们建立了一个基于贝叶斯网络的个人模型的所有者和架构的合作伙伴机器配备了它作为推理模型。个人模型从关注的目标代码、印象词的话语反应和主人周围的上下文推断主人的关注，并推荐什么信息满足他/她的信息需求。通过使用具有基于贝叶斯网络的所有者的个人模型的交互式基于语音的新闻提供者系统的实验，证实了个人模型使得为所有者提供个性化的信息。少

In this research, we proposed fundamental methods for the partner machines such as familiar computers and robots that support daily activities of human owners by paying attention to their daily space and sharing attention and concern with them. These methods include representation and control of synergic attention between an owner and his/her partner machine, learning of attention structure based on the competitive neural network, and Bayesian network-based personal modeling and real-world information navigation using visual attention and response utterance of an owner. As for synergic attention, we built a method for modulating saliency-driven attention by active spatial attention induced by owner's concern. Then we built an attention transition and control method and proposed the synergic attention control method between an owner and his/her partner machine. As for learning of attention structure, we built a model of encoding a set of attended spots in a certain scene as an attention … More structure code. This model consists of the competitive neural network named the COGNET, which encodes objects included in attended spots, and the encoding mechanism of sizes and positions of the attended spots. Through experimental evaluation of the COGNET's main features, that are the fast self-organized learning and the glance recognition of objects in attended spots, it was confirmed that attention structure codes were useful enough for encoding visuospatial episodes. As for personal modeling and information providing based on it, we built a Bayesian network-based personal model of an owner and architecture of a partner machine equipped with it as an inference model. The personal model infers owner's concern from attended object codes, utterance response by impression words and context around the owner, and recommends what information satisfies his/her information needs. Through experiments using an interactive speech-based news provider system with a Bayesian network-based personal model of an owner, it was confirmed that the personal model made information providing personalized for the owner. Less