Weight sharing is a technique in neural networks where the same weights are used across different parts of the network, reducing the number of parameters and improving computational efficiency. This approach is commonly used in convolutional neural networks (CNNs) to enable the network to detect features regardless of their position in the input space.