AI Is Capable Of Solving Physics-Based Problems, New Study Claims

Humans are lousy at solving physics-based problems. We are especially bad when predicting the movement of one or more objects after they interact, and have serious problems when having to calculate the path of a moving object who got launched for another moving object (for instance, the path of a skydiver who jumped from a moving plane). Even with simple tasks, like Conservation task, where water is poured from one short and wide glass to one tall and narrow glass, we can’t tell that both glasses hold the same water amount until we are 7-years old. Physics-based problems are extremely tricky for AI to solve them since they require a wide knowledge of the world and rules in nature, the basic understanding of physical properties of the objects surrounding us, and logical thinking used for connecting the rules.

deep-mind

The AIs weren’t able to solve these types of problems, until recently. Artificial intelligence team behind Google’s Deep Mind AI, assisted by researchers at the University of California, Berkeley, and managed to learn AI machines to evaluate physical properties of an object without any preceding awareness of physical laws.

The study, bearing a title “Learning to perform physics experiments via deep reinforcement learning” used different trials in order to teach the AI how to solve simple physics problems. Two simulations were conducted. The first, titled Which is Heavier, asked the AI to determine which block was heaviest, from the set of four blocks having the same size but different weight. The AI managed to solve the problem via reinforced learning by receiving positive feedback if choosing the heaviest block, and negative feedback when choosing the wrong block. Misha Denil, the head researchers stated that “Assigning masses randomly… ensures it is not possible to solve this task from vision (or features) alone since the color and identity of each block imparts no information about the mass in the current episode.” In other words, the AI learned to determine the weight of the four blocks by learning a new rule that objects sporting the same dimensions don’t have to have the same mass.

Another experiment (titled Towers) featured five blocks that together resembled a tower. The AI had to determine how many blocks were used in the tower. The trick was that some blocks were hidden from vision. The AI had to learn how to find the (on the first look) invisible blocks. It learned to interact with the tower, by pulling it apart in order to learn how many elements built the tower. Instead just passively observing the construction, the AI learned how to interact with the environment in order to reach new conclusions.

deep-mind-tower

Another important milestone in AI research. If scientists manage to make AI learn new things by interacting with the environment, it can develop higher cognitive functions, like reasoning or forward planning, bringing it one step closer to becoming self- conscious. Of course, we are still far away from being able to make AIs with those capabilities, but the experiment showed how an AI can adapt to new rules, and how can reach conclusions without any prior knowledge of the phenomena.

Leave a Reply

Your email address will not be published. Required fields are marked *