Extreme Learning Machines (ELMs), first introduced by Huang et al. (2006), are single-hidden-layer feedforward neural networks where input weights and biases are randomly assigned and remain fixed, while only the output weights are trained, enabling extremely fast learning and strong generalization compared to traditional backpropagation or SVMs. Over the years, numerous variants such as Modified ELM (M-ELM), Weighted ELM (WELM), Constrained ELM (CELM), LARSEN-ELM, kernel-based ELMs, and ensemble approaches have been developed to address challenges like overfitting, robustness, and large-scale learning. Surveys and systematic reviews highlight ELM’s versatility in diverse domains, with applications ranging from image and video processing, document classification, and handwritten digit recognition to medical diagnosis, energy disaggregation, and hyperspectral analysis. Recent research explores multi-layer and convolutional ELMs that bridge the gap with deep learning while preserving the computational efficiency of the original model, establishing ELMs as a fast, scalable, and effective machine learning paradigm with growing applications in both theoretical and practical domains.