基於腦啟發式自我組織情節記憶模型的日常記憶輔助機器人

Abstract

記憶衰退為現代社會常見文明病之一，其中又以老年長者為高罹患族群，時常 伴隨失智症、阿茲海默症等疾病發生，對於社會造成逐漸加重的醫療及照護成本支 出。本論文中提出以一腦啟發式的自我組織情節記憶模型搭配智慧機器人 Pepper ，能於居家環境中透過視覺與聽覺資料接收日常生活事件並提供記憶輔助，其中接 收之事件著重於人、事、時、地、物等五個面向的語義架構用以歸納整理和儲存於 記憶模型之中。本文提出的記憶模型為一啟發於融合式自適應性共振理論(Fusion adaptive resonance theory, Fusion ART)階層式的圖模式系統，針對情節記憶模擬 人腦記憶構造之計算模型，並以一工作記憶暫存區、語意元素(Semantic elements) 層、情節事件層以及事件模板層組成。此系統利用語義指標(Semantic pointers) 的方式編碼記憶中的元素及計算其彼此間的相對距離，同時利用動態時間修正 (Dynamic time warping)和譜分群(Spectral clustering)的方式分類儲存之記憶， 再用詞頻率與逆向文件頻率(TF-IDF)的概念概述各個分群。當需要提供記憶輔助 時，機器人能從使用者的問句之中提取問句目標以及線索之關鍵字詞，並用於分析 及依據其相關面向重新分類已儲存之記憶。同時，透過樸素貝氏分類器(Naïve Bayes classifier)，記憶模型可用關鍵字詞找出具最大機率之最佳符合記憶簇以及其 包含之語意元素，並由機器人以對話形式將顯著語意元素提供給使用者以潛在地 刺激使用者對於該相關事件的再次連結。此外，透過強化學習(Reinforcement learning)，最終機器人能夠從經驗中學習何種記憶提示對於不同的事件類別對於使 用者來說是最有效的。實驗顯示提出之記憶模型的可操作性並同時具備穩定性和 適應性;再者，人機互動的實驗中顯示照護機器人可於此記憶模型建構的知識基礎 上進行穩健的記憶輔助服務，而在接受本研究提出的記憶輔助後，於 99%的信賴 區間下受試者回憶事件的能力有顯著性的提升。

This work shows a household caring robot with a brain-inspired episodic memory model, which provides memory assistance and tackles the modern public issues of memory impairment among individuals and increasing human resource expenditure on elder caring using a humanoid robot, Pepper. The robot is able to visually and auditorily observe user’s daily activities based on five aspects, i.e. people, activities, time, locations and objects, which will then be aggregated and stored as episodes in the self-organizing memory model. The proposed memory model is a hierarchical graph-based system inspired by fusion adaptive resonance theory (Fusion ART) simulating a computational model of the human brain targeting episodic memories, comprising a working memory buffer, semantic element layer, episodic event layer, and event class layer. The system uses semantic pointers to encode event elements and calculate the relative distances among them, while a modified spectral clustering with dynamic time warping is implemented to merge and categorize the observed memories, where clusters are summarized with the concept of Term Frequency - Inverse Document Frequency (TF-IDF). When providing memory assistance, the robot is able to spot the keywords of the query target and the hints in the query issued from the user, which are later used for analysis and memory re-classification in related aspects. Furthermore, with naïve Bayes classifier, the model finds the best matching memory cluster and its semantic elements with the highest probability associated with the keywords in the query. Next, accordingly, the robot provides the memory cue with the observed salient semantic elements regarded as context to the user through dialog, which potentially may stimulate the user’s recall of the target event related to the query. Besides, using reinforcement learning, the robot eventually learns what kinds of memory cues are the most effective supports to the user for each type of event from experience. The experiments show the feasibility of the proposed model, which could handle episodic events with elasticity and stability. Moreover, from the HRI experiment, the robot is able to provide robust memory assistance from the knowledge of the memory model, with 99% confidence intervals the participants’ mean recall percentage of the events increases significantly after receiving the proposed memory assistance.