On Error Probability and Computational Complexity of Pattern Recognition in a Metric Space of the Tree-Structured Representations.

Autor: Lange, M. M., Paramonov, S. V.
Zdroj: Optoelectronics Instrumentation & Data Processing; Oct2022, Vol. 58 Issue 5, p440-447, 8p
Abstrakt: We study a classification accuracy in terms of dependence of an error probability on an amount of processed information in a space of the tree-structured representations of recognized objects. For a specified set of the objects, a lower bound to the error probability as a function of an average mutual information between the objects and their class label decisions is given. Using the multilevel discriminant functions, an algorithm is proposed for guided search for an object decision, and the computational profit of this algorithm is estimated as compared to exhaustive search. The experimental dependences of the average error probability and average mutual information on the parameter characterizing the computational profit of this algorithm are demonstrated for the image-based datasets of faces and signatures and the ensemble of these datasets. The lower error probability boundaries, which make it possible to estimate the algorithm error probability redundancy at different computational profit values, are calculated for the above mentioned data sources. [ABSTRACT FROM AUTHOR]
Databáze: Complementary Index
Externí odkaz: