In his recently published paper, Kaiser outlined how a system guided by a decision-making engine of sorts can learn how to play competently games with only a minimal amount of background data.

This is where things get a little technical, so bear with us: Kaiser states that while computer scientists have done a great amount of work in regards to computerized object recognition and visual scene interpretations, "only a few systems with the capacity for learning higher-level concepts has been presented thus far." According to Kaiser, our computers are pretty good at deriving sequences of higher-level symbolic data from video streams, but we still have a long way to go when it comes to learning from it.

He argues that a more nuanced approach using relational structures and multiple logic systems is better suited for learning from visual data in comparison to the standard practice of utilizing formulas and singular logic systems. "These two fundamental changes allow us to demonstrate a system that--knowing only about rows, columns, diagonals and differentiating pieces--learns games like Connect Four, Gomoku, Pawns or Breakthrough, each one from a few intuitive video demonstrations, together around 2 minutes in length."

Is this where we start preparing for the rise of the machines? Not quite yet. Kaiser still needs to figure out how to get the system to solve problems requiring "hierarchical, structured learning or a form of probabilistic formulas." Until then, we're safe. After that, it's anyone's game.

[ via ]