I'm surprised they refer to their technique as a "growing neural gas", as it appears to be mostly an on-line graph-based manifold learning technique. See ex. Laplacian eigenmaps: <a href="https://en.wikipedia.org/wiki/Nonlinear_dimensionality_reduction" rel="nofollow">https://en.wikipedia.org/wiki/Nonlinear_dimensionality_reduc...</a>