"Tell me and I forget. Teach me and I remember. Involve me and I learn." Benjamin Franklin
Knowledge Until very recently, most "learning" or knowledge was based upon experimentation and observation. This is how learning and discovery happened. A person became an expert by experimentation, observation and learning from other experts. Machine Learning (ML) allows us to learn or gather knowledge from data itself. With ML knowledge is gathered in an automated fashion using computer based algorithms. These algorithms try to discover underlying patterns in the data.
Not Explicitly Programmed Imagine teaching a computer to play chess. One way is a programmer programs all the rules. A programmer could study a rule book on the rules of chess. The programmer could explicitly program the rules such as a rook moves forward and backwards, a bishop moves diagonal, so on and so forth. Now imagine instead of explicitly programming all the rules a computer analyzes all the chess moves from millions of chess games. In the second scenario the computer "learns" the rules for the game of chess.
How Does A Computer Learn? A computer learns patterns. A data scientist writes algorithms so the machine (the computer) can learn. The data scientist has to tell the "machine" which algorithms to use to find patterns. Within the branch of ML there are thousands of automated techniques to help understand and discover patterns.
Learning the Wrong Thing - Learn Verify Train There is a great scene in the movie Starman. Starman, played by Jeff Bridges, learns to drive a car by observing Jenny Hayden (Karen Allen). In the scene Jenny becomes terrified as Starman races through an intersection. She says, "You said you knew the rules!" Starman responses with "I watched you very carefully, Red Light Stop, Green Light Go, Yellow Light Go Very Fast." (video below). Within all data there are biases, the wrong thing. An inherent problem with Machine Learning is the wrong thing can be learned. It is problematic when the wrong thing becomes part of the set of learned rules. It would be analogous to teaching a computer a bad habit.
If "learned rules" are going to be productionalized a necessary step is to verify. In other words, teach the computer the right thing not the wrong thing.