In theory, each head in multi-headed attention learns something different, therefore giving the encoder model more representation power through multiple perspectives on the same input.