Self-Attention is a mechanism where each position in a sequence calculates attention weights for every other position. It helps models capture long-range dependencies without recurrent or convolutional operations. Transformers rely heavily on self-attention for state-of-the-art NLP results.