How can we define Machine Leanning by itself?

Suppose you are learning to recognize faces. A completely empty system has to be told what parts of images are faces. The more faces the system presents in which it has been told which parts are faces the more likely the system will get it right. It isn’t unlike a spam filter where it needs samples of what are spam and what are not before it can start seeing similarities in random cases without a human to help. Once enough samples are given in which someone tells the system the answers (the training) the system may become faster and better than a human at the same thing (think hundreds of thousands of images). It learns by being told what is right or wrong…it can’t know that without help.