loading1
loading2

Small Dynamic Neural Networks for Gesture Classification with The Rulers (a Digital Musical Instrument)

category

author

John Sullivan

Download available

Authors:

Vanessa Yaremchuk, Carolina B. Medeiros, Marcelo M. Wanderley

Publication or Conference Title:

Proceedings of the 2019 International Conference on New Interfaces for Musical Expression (NIME 2019)

Abstract:

The Rulers is a Digital Musical Instrument with 7 metal beams, each of which is fixed at one end. It uses infrared sensors, Hall sensors, and strain gauges to estimate deflection. These sensors each perform better or worse depend- ing on the class of gesture the user is making, motivating sensor fusion practices. Residuals between Kalman filters and sensor output are calculated and used as input to a re- current neural network which outputs a classification that determines which processing parameters and sensor measurements are employed. Multiple instances (30) of layer recurrent neural networks with a single hidden layer vary- ing in size from 1 to 10 processing units were trained, and tested on previously unseen data. The best performing neu- ral network has only 3 hidden units and has a sufficiently low error rate to be good candidate for gesture classification.
This paper demonstrates that: dynamic networks out- perform feedforward networks for this type of gesture classification, a small network can handle a problem of this level of complexity, recurrent networks of this size are fast enough for real-time applications of this type, and the importance of training multiple instances of each network architecture and selecting the best performing one from within that set.


Publication Details:

Type:
Conference Paper
Date:
06/01/2019
Pages:
150-155
Location:
Porto Alegre, Brazil
DOI:
10.5281/zenodo.3672904

IDMIL Participants:


Related Projects