previous, up, next

General Approach

Growing self-organizing networks are vector-based models trained by competitive learning and also (in case of the supervised variants of the models) by simple supervised techniques such as the Delta rule. The general strategy is to start with a near-minimal network and add units until the network has the desired size or is "good enough" for the task at hand. This approach can be characterized by:

• generation of the network structure through a growth process
• use of local statistical quantities to guide insertions and deletions of units

Different constraints imposed on the network structure lead to different models:

• general graph: growing neural gas
• k-dimensional hypertetrahedrons: growing cell structures
• rectangular grid: growing grid

Advantages over other approaches include:

• only constant parameters
• network structure and size has not to be defined in advance
• fast training, in particular in supervised learning (a factor of 30-100 over backprop has been observed for some problems).

(Last updated: June 24, 1997)