Self-organizing networks such as neural gas, growing neural gas and many others have been adopted in actual applications for both dimensionality reduction and manifold learning. Typically, the goal in these applications is obtaining a good estimate of the topology of a completely unknown subspace that can be explored only through an unordered sample of input data points. In the approach presented here, the dimension of the input manifold is assumed to be known in advance. This prior assumption can be harnessed in the design of a new, growing self-organizing network that can adapt itself in a way that, under specific conditions, will guarantee the effective and stable recovery of the exact topological structure of the input manifold.