‘Machine learning, especially reinforcement learning, needs an objective function, also called a loss, utility, or fitness function. This is the equation that defines what counts as good or bad, and the system learns by optimizing it toward a clear goal, like winning in chess. That is what makes machine learning different from evolution. In evolution, the “goal” seems to be adaptation or survival, but what counts as the “fittest” is often not clearly defined in advance. It emerges through interaction with the environment.
In machine learning, that goal does not emerge on its own. Human ingenuity provides it. Humans define the objective function, and that function determines which system performs better. The deeper point is that humans also seem to search for their own objective functions. We ask what our purpose is and why we are here. Machines, by contrast, do not do that on their own. At least for now, they need a hard-coded statement of what their “why” is.’
-Thoughts from Dr Lex Fridman on The Huberman Lab podcast. He is a Ph.D & scientist at MIT (Massachusetts Institute of Technology), working on robotics, artificial intelligence, autonomous vehicles and human-robot interactions.
