We develop a theory of neural representation that considers a neural s
ystem as an autonomous entity that adapts its internal structure on th
e basis of subjective probabilities constructed on the basis of random
ly received input signals. The objective probabilities distribution of
the signal space is principally unknown to the system, and furthermor
e the system's limitations, e.g. structural rigidity and limited capac
ity form an integral part of our analysis. These considerations yield
the conceptual basis for a new algorithm for a self-organizing feature
map.