Akademik Veri Yönetim Sistemi
6th International Applied Statistics Congress, Ankara, Türkiye, 14 - 16 Mayıs 2025, ss.602-609, (Tam Metin Bildiri)
In this study, we propose SoftReMish, a new activation function designed to improve the performanceof convolutional neural networks (CNNs) in image classification tasks. Using the MNIST dataset, weimplemented a standard CNN architecture consisting of two convolutional layers, max pooling, andfully connected layers. We evaluated SoftReMish against popular activation functions including ReLU,Tanh, and Mish by replacing the activation function in all trainable layers. The model performancewas assessed in terms of minimum training loss and maximum validation accuracy. Results showedthat SoftReMish achieved a minimum loss(3.14 × 10⁻⁸) and a validation accuracy(%99.41)outperforming all other functions tested. These findings demonstrate that SoftReMish offers betterconvergence behavior and generalization capability, making it a promising candidate for visualrecognition tasks.