NELE: Non-Uniform Rational B-Spline Elementwise Learnable

This work introduces an activation function (AF) based on Non-Uniform Rational B-Splines (NURBS), designed to provide smooth, flexible, and learnable non-linearities for neural networks. Unlike traditional fixed AF (ReLU, tanh, sigmoid), the NURBS-based activation can adapt its shape during training, offering higher expressivity while maintaining differentiability. Alternatively, the NURBS curve can also be used to represent traditional AF. An archive containing trained PyTorch model weights (.pth) for NELE activation function on the NLSD, MNIST, MNIST encoder, CIFAR10 and CIFAR100 dataset is provided. These weights correspond to the results reported in the associated work and can be used for inference, evaluation, or further research. Loss vs epoch trends are provided as .csv file