Neural Networks have shown terrific results in working out several tasks accurately. Be it speech recognition, computer vision, natural language processing, or even generative intelligence, Artificial Neural Networks are ubiquitous. These clusters of neurons with intermittent connectivity are also referred to as Multilayered Perceptron (MLP). Since its inception in 1958, MLPs have served a lot to the neural networking world. It consists of a cluster of neurons assorted at different levels in the form of layers that are connected using non-linear fixed activation functions. These neurons mimic the learning capabilities of the human brain, as it learns the weights and biases annotated with the respective neurons by the process of Backpropagation. Recently in March 2024, inspired by the Kolmogorov-Arnold representation theorem, Liu et al. proposed the Kolmogorov-Arnold Networks (KANs), which grew out of an effort to have learnable, activation functions in place of the weights of an MLP. KANs possess no linear weight vectors, rather it replaces the entire weights as evident in the MLPs with univariate activation function (here., Spline). In the introductory dissemination of KANs, the authors showed better accuracy attained by the KANs over MLPs for a wide range of tasks. Following the introductory article, several advancements have been contributed in the same direction like TKAN, Wav-KAN, DeepOKAN, etc. This review aims to study these recent developments and their respective prospects. Each of these improvisations will be thoroughly examined based on their target domain and scopes of further improvisations.
