Learning constituent parts of touch stimuli from whole hand vibrations
2016 IEEE Haptics Symposium (HAPTICS)(2016)
摘要
Manual activities elicit touch stimuli that are widely distributed in the hand. To what extent can these stimuli be considered to the sum of their parts, and how might these parts relate to the structure and function of the hand? In this work, we measured spatially distributed patterns of vibration in the skin that were elicited during common manual interactions. We analyzed information content in this data by extracting informative low-dimensional representations, using an unsupervised machine learning algorithm known as non-negative matrix factorization. The latter is a technique for decomposing data into implicit components, inspired by neural processing. This analysis automatically yielded "parts of touch" - cohesive representations of touch-elicited vibration patterns that were localized in the fingers and other cohesive areas of the hand, reflecting to a remarkable extent, the most salient anatomical and functional specializations in the hand. In a subsequent classification task, we found these representations efficiently encoded the manual activities that elicited the (high-dimensional) vibration stimuli. The activities could be accurately classified using low dimensional representations with as few as three parts. The results provide quantitative evidence that vibrotactile information content in the upper limb is organized in ways that reflect the anatomy, structure, and function of the hand.
更多查看译文
关键词
touch stimuli,whole hand vibrations,manual activities,spatially distributed vibration patterns,informative low-dimensional representations,unsupervised machine learning algorithm,nonnegative matrix factorization,neural processing,touch-elicited vibration patterns,functional specializations,salient anatomical specializations,quantitative evidence,vibrotactile information content
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络