Two of the most commonly used pooling methods in Convolutional Neural Networks are maximum and average pooling. The maximum pooling method may lead to the loss or reduction of discriminative features in high-intensity inputs. On the other hand, in the average pooling method, when average values calculated from local regions are of low intensity, it can cause a decrease in the contrast of feature maps. Therefore, it has been observed that the hybrid pooling method, which weights and combines the outputs of both methods, performs better in some applications. In this study, it is demonstrated that the optimal value of the learnable parameter used for weighting may not be found when updated with all model parameters during the training process. To address this issue, the weighting in hybrid pooling is optimized independently from the model using a Genetic Algorithm (GA). In a multi-classification study conducted with the Flowers dataset, the performance of multi-classification obtained through GA optimization was found to be higher compared to the traditionally learned weight model, based on test results. The accuracy values of the model optimized with GA and the models containing traditionally learned weight were found to be 0.8299 and 0.8109, respectively.
Paper
Full text
Optimization of Mixed Pooling Using Genetic Algorithm for Convolutional Neural Networks
Semantic Scholar · Computer Science · 2024
Abstract
Two of the most commonly used pooling methods in Convolutional Neural Networks are maximum and average pooling. The maximum pooling method may lead to the loss or reduction of discriminative features in high-intensity inputs. On the other hand, in the average pooling method, when average values calculated from local regions are of low intensity, it can cause a decrease in the contrast of feature maps. Therefore, it has been observed that the hybrid pooling method, which weights and combines the outputs of both methods, performs better in some applications. In this study, it is demonstrated that the optimal value of the learnable parameter used for weighting may not be found when updated with all model parameters during the training process. To address this issue, the weighting in hybrid pooling is optimized independently from the model using a Genetic Algorithm (GA). In a multi-classification study conducted with the Flowers dataset, the performance of multi-classification obtained through GA optimization was found to be higher compared to the traditionally learned weight model, based on test results. The accuracy values of the model optimized with GA and the models containing traditionally learned weight were found to be 0.8299 and 0.8109, respectively.