Growing neural gas

Growing neural gas (GNG) is an artificial neural network algorithm that learns the topology of input data by incrementally adding neurons and connections. Unlike the self-organizing map, which uses a fixed grid, GNG grows its structure dynamically: neurons are inserted near the neuron with the highest accumulated error, and connections are created between the best-matching unit and its neighbors.

The algorithm was developed by Bernd Fritzke in 1995. It is particularly useful for learning non-stationary distributions and for tasks that require an adaptive representation of the input space, such as robotics motion planning and computer vision feature extraction. GNG can be seen as a bridge between neural network topology learning and computational geometry mesh generation.