He Initialization is a weight initialization strategy designed for ReLU or variants. It sets layer weights with a variance of 2/n, where n is the number of input units in that layer. This helps maintain stable gradients and fosters efficient training in deep networks.