What is Machine Learning?
Machine Learning is when a computer runs an algorithm that is capable of self improvement. For instance, a computer learns to create images of dogs by first randomly placing pixels, then comparing that image to a database of images of dogs. The computer will eventually create an image that is somewhat similar to a dog, and it will be rewarded. From there, it will slowly improve until it can create its own images of dogs.
What if you have no pre-made input data?
For some problems, a preexisting data source is needed. Usually, any image or text based algorithms require data, or else they will never make sense to a human.
However, there are many problems that can be solved with no human readable data. One such example is my project, the Hummingbird AI. It works by receiving a bunch of numbers: its pitch, yaw, rotation, beak position, body position, speed, location, and proximity to the nearest flower, and seven rays that tell the AI how far it is from objects nearby. It then outputs numbers between -1 and 1 for horizontal, vertical, and rotational movement. When the AI first starts training, it sends outputs randomly, but over time, it learns what series of inputs and outputs lead to a higher reward, and so it does those things. It has no concept of a flower, or a hummingbird, or nectar. All it knows is that if a number is remarkably close to zero and another number is rising (the beak is in a flower), then it gets a reward. It doesn't even "see" another flower until it finishes with its current target!
Another, far more complex example is OpenAI's Hide and Seek Simulation. I am not educated enough to know exactly how they did it, but to summarize, two AIs were developed: hiders and seekers. These two AIs would be pitted against one another and over time, they both learned and developed new strategies. They even learned ways to break the rules of the simulation! If you want to learn more, check out OpenAi.