Localized Learning: A Possible Alternative to Current Deep Learning Techniques

Authors

Javier Viana, University of CincinnatiFollow
Kelly Cohen, University of CincinnatiFollow
Anca Ralescu, University of CincinnatiFollow
Stephan Ralescu, University of CincinnatiFollow
Vladik Kreinovich, The University of Texas at El PasoFollow

Publication Date

9-1-2021

Comments

Technical Report: UTEP-CS-21-85

Abstract

At present, the most efficient deep learning technique is the use of deep neural networks. However, recent empirical results show that in some situations, it is even more efficient to use "localized" learning -- i.e., to divide the domain of inputs into sub-domains, learn the desired dependence separately on each sub-domain, and then "smooth" the resulting dependencies into a single algorithm. In this paper, we provide theoretical explanation for these empirical successes.