Despite numerous studies on pooling algorithms, max-pooling remains the standard choice in CNNs. In this work, we introduce (a, b)-grouping functions, an extension of grouping functions tailored for real-valued data in CNNs. We present various construction methods for (a, b)-grouping functions and empirically demonstrate their effectiveness by replacing max-pooling in popular CNN architectures, yielding promising results. Our findings highlight the potential of grouping functions as efficient alternatives to max-pooling in CNN feature downsampling.
Iosu Rodríguez Martínez, Tiago Asmus, Graçaliz Dimuro, Francisco Herrera, Zdenko Takáč, Humberto Bustince