Low‐Dimensional Embeddings for Interaction Design. (27th July 2021)

• Record Type: Journal Article
• Authors: Rusu, Marius Mihai; Schött, Svenja Yvonne; Williamson, John H.; Schmidt, Albrecht; Murray-Smith, Roderick
• Is Part Of: Advanced intelligent systems. Volume 4:Number 2(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-07-27
• Subjects: autoencoders -- data gloves -- dimensionality reduction -- embedding -- interaction; Artificial intelligence -- Periodicals; Robotics -- Periodicals; Control theory -- Periodicals; 006.3
• DOI: 10.1002/aisy.202100045
• Languages: English
• ISSNs: 2640-4567

Abstract

Abstract : Physical interactions with the real world have many degrees of freedom, which has led to the development of novel input devices with a multitude of sensors to capture increasingly high‐dimensional data. This high dimensionality makes the design of interactive systems more complex. Herein, the use of autoencoder‐based dimensionality reduction is explored to simplify the design process. For this purpose, a data glove equipped with accelerometers is used to record high‐dimensional hand movement data that are thereafter reduced to 2D embeddings using autoencoders. The exploration and evaluation of the resulting embeddings suggest that autoencoders can be used to create meaningful low‐dimensional representations of complex human movement. The characteristics generality, variability, connectivity, and distinguishability are established and a guideline is provided for assessing low‐dimensional embeddings. Referring to these characteristics, system engineers can evaluate different input modalities and gestures for their specific interaction task. Further, a framework is outlined for designing and evaluating gesture interaction in the low‐dimensional space. By demonstrating the exemplary design of the interaction with a virtual lever, this research gives system engineers a template for interaction design in the low‐dimensional space. Abstract : How can interaction design be improved? How can the complexity new sensors provide be kept up with? Dimensionality reduction, a form of machine learning, is a viable option. Low‐dimensional representations are created from high‐dimensional data, which can be used to design and analyze interaction tasks in a comprehensible way.