GaussBox

May 6, 2016 #HMM #Visualization

GaussBox is a pedagogical tool for prototyping movement interaction using machine learning. GaussBox proposes novel, interactive visualizations that expose the behavior and internal values of probabilistic models rather than their sole results. Such visualizations have both pedagogical and creative potentials to guide users in the exploration, experience and craft of machine learning for interaction design.

Principle

Gaussbox-principle

Demonstration

Software

GaussBox will be available **sometime** as an open-source application. Gaussbox supports several types of input devices (mouse, Leap Motion, Myo, OSC). It allows for playing with both real-time recognition and continuous mapping to sound (with CataRT-style sound synthesis or OSC).

Reference

  • Jules Françoise, Frédéric Bevilacqua, and Thecla Schiphorst, “GaussBox: Prototyping Movement Interaction with Interactive Visualizations of Machine Learning,” in Proceedings of the 2016 CHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA '16), San Jose, CA, ACM, 2016, pp. 3667--3670. DOI: 10.1145/2851581.2890257. http://dl.acm.org/authorize?N03765.
    Abstract
    We present GaussBox, a design support tool for prototyping movement interaction using machine learning. In particular, we propose novel, interactive visualizations that expose the behavior and internal values of machine learning models rather than their sole results. Such visualizations have both pedagogical and creative potentials to guide users in the exploration, experience and craft of machine learning for interaction design.
    Download Project Page
    Acceptance Rate: 20%
  • Jules Françoise, Thecla Schiphorst, and Frédéric Bevilacqua, “Supporting User Interaction with Machine Learning through Interactive Visualizations,” in CHI'16 Workshop on Human-Centred Machine Learning, San Jose, CA, 2016. http://www.doc.gold.ac.uk/~mas02mg/HCML2016/HCML2016_paper_20.pdf.
    Abstract
    This paper discusses novel visualizations that expose the behavior and internal values of machine learning models rather than their sole results. Interactive visualizations have the potential to shift the perception of machine learning models from black-box processes to transparent artifacts that can be experienced and crafted. We discuss how they can reveal the affordances of different techniques, and how they could lead to a deeper understanding of the underlying algorithms. We describe a proof-of-concept application to visualize and manipulate Hidden Markov Models, that provides a ground for a broader discussion on the potentials and challenges of interactive visualizations in human-centered machine learning.
    Download Project Page
Featured image