Thought experiment.

I'm pretty much a novice in machine learning, but this thought experiment occurred to me:

What if you had worn a gps recorder which records every 5 minutes, and is accurate to say 2 cm for the last 10 years, and you were going to use the last 10 years of data to predict your entire pattern of movement for the next day. The most accurate predictor of this, would require a rule system which encodes your entire brain and its functions, and the entire environment that you interact with, including other people, and the laws of physics. In fact, to be as accurate as possible it would have to simulate the entire world. What if you turn on the news and you see a fire is near your house, and then you drive away? Well if you had included the wheather patterns and simulated forests and the likelyhood of forest fires you would have been more likely to predict the pattern of movement.

So this would require ridiculously massive amount of computation.

So, when something that seems as simple as your gps coordinates turns out to require computing an whole world, how accurate can machine learning possibly be for many things? Maybe this explains why stuff like facial recognition is so difficult. Because to calculate the likelyhood of a face, you have to be able to take an image and take the 2 dimensional pixels and convert it to a 3d surface, which in order to solve correctly would require advanced knowledge of how our world works, including even the type of camera used, the exact location of the camera in 3d space relative to the face, and probably the genetics of how human heads are shaped, including ethnicity, skin tone, ect. Its a ridiculously complicated problem, so to think that a machine learning algorithm will figure all that out is insane.