Synthetic pattern recognition methods based on the representational minimum description length principle
Alexey S. Potapov · 2008
The minimum description length principle (MDL) is considered in context of pattern recognition problem. It is shown that existent applications of the MDL principle use heuristic coding schemes in order to calculate code length in contrast to the theoretic MDL principle based on the notion of algorithmic complexity. The representational MDL principle is proposed as an extension of the MDL principle that replaces heuristic schemes with the notion of information representation with strict measure of quality. Pattern recognition problem is reduced to the task of choosing such the optimal representation that minimizes summed description length of patterns in a training set. As a result, synthetic methods with automatic selection between particular methods are proposed.