Identity activation structural tolerance online sequential circular extreme learning machine for highly dimensional data
Sarutte Atsawaraungsuk, Tatpong Katanyukul, Pattarawit Polpinit · DOAJ (DOAJ: Directory of Open Access Journals) · 2019
The Structural Tolerance Online Sequential Circular Extreme Learning Machine (STOS-CELM) was developed based on the Circular Extreme Learning Machine (CELM) to allow sequential learning and to mitigate the criticality of deciding the number of hidden nodes withtheHouseholder Block exact inverse QRD Recursive Least Squares (HBQRD-RLS) algorithm. Aprevious study showedsignificant efficiency improvement using STOS-CELM with sine activation. However, sine activation is periodic. Its periodicity repeatedly maps multiple values of its input to the same output values. Within the context of STOS-CELM, input of the activation is a non-negative real value corresponding to the closeness of the data to CELM kernels. Mapping this non-negative real value to a limited range witha periodic nature causes loss of inherent information. That could restrain the STOS-CELM from reaching its full potential. This article proposes an Identity Activation Structural Tolerance Online Sequential Circular Extreme Learning Machine (ISTOS-CELM) to improve the STOS-CELM by removingthe sine function to relievethisissue. Our experimental results show that ISTOS-CELM provides significantly higher accuracy than STOS-ELM and the original STOS-CELM, while retaininga comparable processing time and robustness to STOS-CELM.