TIGP (SNHCC) -- Evolutionary Quadtree Pooling for Convolutional Neural Networks
- LecturerProf. Po-wei Harn (Department of Computer Science & Information Engineering, National Central University.)
Host: TIGP (SNHCC) - Time2026-03-02 (Mon.) 14:00 ~ 16:00
- LocationAuditorium 106 at IIS new Building
Abstract
Despite the success of Convolutional Neural Networks (CNNs) in computer vision, it can be beneficial to reduce parameters, increase computational efficiency, and regulate overfitting. One such reduction technique is the use of so-called pooling, which gradually reduces the spatial dimensions of the data throughout the network. Recently, Quadtree-based Genetic Programming has achieved state-of-the-art results for optimizing spatial areas on customized requirements in different grid structures.
Motivated by its success, we propose to extend this approach to pooling layers of CNNs. In this direction, this paper introduces a new way to look at each pooling layer. Specifically, we propose an Evolutionary Quadtree Pooling (EQP) method that can identify the best pooling scheme. By embedding multiple quadtrees set as a pooling scheme in the pooling layers of a CNN, we are able to operate crossover and mutation on the feature maps. The evolutionary process of EQP guides the search to provide more reliable evaluations, where each individual can be seen as a CNN with a new type of pooling scheme. Our experimental results show that the best candidate network of EQP outperforms state-of-the-art max, average, stochastic, median, soft, and mixed pooling in accuracy and overfitting reduction while maintaining low computational costs.
Motivated by its success, we propose to extend this approach to pooling layers of CNNs. In this direction, this paper introduces a new way to look at each pooling layer. Specifically, we propose an Evolutionary Quadtree Pooling (EQP) method that can identify the best pooling scheme. By embedding multiple quadtrees set as a pooling scheme in the pooling layers of a CNN, we are able to operate crossover and mutation on the feature maps. The evolutionary process of EQP guides the search to provide more reliable evaluations, where each individual can be seen as a CNN with a new type of pooling scheme. Our experimental results show that the best candidate network of EQP outperforms state-of-the-art max, average, stochastic, median, soft, and mixed pooling in accuracy and overfitting reduction while maintaining low computational costs.