Self Organizing maps Vs k-means

Question

Does anyone know how well does Self Organizing Maps(SOM) compare to k-means? I believe usually in the color space,such as RGB, SOM is a better method to cluster colors together as there is overlap in the color space between visually different colors (http://www.ai-junkie.com/ann/som/som1.html). Are there cases where k-means outperforms SOM?

Thanks!

Hewart

Answer 1

K-means is a specialisation of SOM, I believe. You can construct ideal cases for it, I'm sure. I think computational speed is its major advantage -- when you have incrementally improving AI algorithms, sometimes more iterations of a worse algorithm gives better performance than fewer iterations of a bettwer, slower algorithm.

It all depends on the data. You never know until you run it.

Pastrami

Answer 2

K-means is a subset of Self-Organizing Maps (SOM). K-means is strictly an average n-dimensional vector of the n-space neighbors. SOM is similar but the idea is to make a candidate vector closer to the matching vector and increase the difference with surrounding vectors by perturbing them; the perturbation decreases (kernel width) with distance; that is where the Self-Organizing part of the name comes from.

Horme

this doesn't sounds right to me. I would consider SOM as a more restrictive version of K-means (not the other way around). – Burnet

Answer 3

Self Organizing Maps create a 2-dimensional output. k-means is multi-dimensional. SOMs operate in a discretized representation (grid). SOMs use a more local rule (neighborhood function). k-means is more widely used as a clustering algorithm.

Zsa

SOMs do not necessarily create 2D output. Most implementations do, but there also exist SOMs with 3D lattices. SOMs usually map high dimensional codebook vectors onto lower-dimensional vectors in a lower-dimensional lattice, which is not limited to 2D, 3D, etc by any means. – Dillard