Particle Trajectories in a Data Potential

This sonification model allows to get an auditory presentation of high-dimensional data using an deterministic dynamic process. Given an arbitrary high-dimensional function V(x), test particles are injected in the domain of V. Regarding V as an potential function and using Newtons laws of motion, these test particles move along a well defined path or trajectory through data space. Given a data set {x_i}, i=1..N, V can be constructed by a superposition of 1-point potentials phi(x-x_i) which are shifted to data point x_i. Using negative Gaussians with bandwidth sigma, a smooth function V is the result.
The test particles are moving in data space according to Newtons law of motion. They are considered as point masses with a given constant mass m and initial Energy E_0. However, we added a friction term to the equations of motion so that the particles converge to local minima of V. These potential troughs correspond to clusters in the data.

No, consider, we throw 50 test particles into data space at random position, compute their trajectories and use the kinetic energy as a function of time as the sound pressure. What do we expect to get ? The particles will move around the domain of V until their energy loss forces them to move into throughs of V. The deeper they get into the trough, the more harmonic is the shape of V. We know that particles in an harmonic potential will perform quasiperiodic motions. Thus, the kinetic energy of the particles will also vary periodically. This will be heard as pitched tones.
 

Literature

For further details on this model, take a look to  Listen to your Data: Model-Based Sonification of Data Analysis

Contact

Thomas Hermann:  thermann@techfak.uni-bielefeld.de

Back to Sonification Page



Last modified: 2001-05-22