next up previous contents
Next: 3.2.1 Algorithm Up: 3. Mixture of local Previous: 3.1 Motivation for local

3.2 Extension of Neural Gas to local PCA

The vector-quantization algorithm Neural Gas (section 2.2.4) is extended such that each code-book vector includes a local PCA (Hoffmann and Möller, 2003; Möller and Hoffmann, 2004). Henceforth, the new algorithm is called NGPCA. We chose Neural Gas because it results in a near homogeneous coverage of data distributions independent of the initial position of code-book vectors. First, NGPCA is described; then, a modification about the competition between units is presented. In the last part of this section, simulations demonstrate the operation of the algorithm.


Heiko Hoffmann