Extreme Learning Machines (ELMs) refer to a category of feedforward neural networks for classification, regression, clustering, sparse approximation, and feature learning. Unlike conventional feedforward neural networks, ELMs randomly assign weights to the connections between the input layer and the hidden layer, and then analytically determine the output weights.

Here’s a brief overview:

1. Random Initialization: The weights and biases between the input layer and the hidden layer are randomly initialized. This means they are not altered during training, which reduces the computational cost.
2. Activation Function: The hidden layer neurons apply an activation function (like the sigmoid function) to the linear combination of the input data and weights. This introduces nonlinearity into the model.
3. Analytical Determination of Output Weights: The weights between the hidden layer and the output layer are determined analytically using techniques like Moore-Penrose generalized inverse. Since this is an exact solution, it avoids iterative tuning and can be very fast.
4. Single-Layer Training: The simplicity of training only one layer makes ELMs faster than traditional neural networks, which may require extensive tuning of multiple layers’ weights through backpropagation.
5. Generalization Performance: ELMs are often effective in terms of generalization performance, meaning they can perform well on unseen data.
6. Limitations: The random initialization can lead to variability in performance. Additionally, ELMs might not be suitable for all problems, particularly complex ones where careful tuning of all the weights is necessary.

ELMs are often used when a quick solution is needed, and where the increased speed and simplicity of the model are more important than achieving the absolute highest accuracy. They have been applied to various domains, including pattern recognition, regression analysis, and classification tasks.