Introduction to Bayesian Inference
Vector Quantization
> SIRIEN Home > Cours Apprentissage Automatique > Demos > Vector Quantization

Systèmes Interactifs et Robotiques

Vector Quantization
  by Frezza-Buet Herve

Growing Neural Gas with Targetting (GNG-T)

Structure extraction

GNG-T enables to find out the structure of non-stationary distributions.

The labelling of connected components adds more structural information.

Stabilization issues

Controlling the number of vertices directly with the target raises stabilization issues.

The use of a smallest confident interval analysis improves stability.

The aperture problem

GNG-T adaptability favors the aperture problem.

A blocking of GNG-T graph updates (i.e. a transiant SOM behavior) helps.

Velocity field computation