Activation functions are essential in the overall performance of deep learning models as they alter learning efficiency and generalization ability. In this work, we propose a new piecewise combination approach to constructing activation functions using existing functions with different activation functions for negative input regions. To experimentally evaluate the proposed method on four publicly available datasets: Iris Flower, Breast Cancer Wisconsin (Diagnostic), MNIST, and CIFAR10. Our findings indicate that different datasets perform best with different combinations of activation functions. For the Iris Flower dataset, the best accuracy is achieved by using ReLU and Sigmoid together. In contrast, the Breast Cancer Wisconsin (Diagnostic) dataset reaches its highest accuracy with ReLU and Mish. A combination of ReLU and Swish works best for MNIST and CIFAR10 in the 4-CNN model, while MNIST and CIFAR10 perform better with ReLU and Tanh in VGG16. Custom piecewise activation functions can improve how well neural networks work with different datasets. This presents a new way to create better activation functions for deep learning models.