Google DeepMind researchers trained the AI system to learn about the various physical properties of objects using two different virtual environments. Thus, they created two different scenarios which produced same results.
The first scenario had five blocks of the same size assigned to the AI. However, the blocks had different mass each time the experiment ran. The AI was provided with a positive feedback if it correctly identified the heaviest block and a negative feedback if it was wrong.
Through trial and error without any instructions, the AI figured out how to achieve the correct results. It learned that the only way to determine the heaviest block was to interact with all of them before making a choice.
The second scenario was similar to the first one. It also featured five blocks arranged in a tower. This time, some blocks were a part of a larger block while others were not. The AI had to find out how many distinct blocks were there , again receiving feedback depending on its answer. In the end, the AI learned it had to interact with the tower, dismantling it in the process, to determine the correct number.
A familiar technique
This technique where computers are training through rewards and punishment is called deep reinforcement learning. It allows solving different tasks without any prior instructions, similar how humans or animals are able to solve problems.
DeepMind is a rather popular figure in terms of using deep reinforcement learning. In fact, it used the method several times to beat humans in both physical and virtual games.
The advancements in teaching AI about the physical properties of the environment will be especially relevant in robotics. This will help them coordinate in the real world and navigate around potential obstacles.
DeepMind is a neural network created by DeepMind Technologies. Google acquired the company in 2014 and renamed it Google DeepMind. It is a member of “Partnership on AI”, an organization whose aim is using artificial intelligence to improve society, along with Amazon, Facebook, IBM, and Microsoft.
The members of the organization will contribute research papers with an open license in many subjects including ethics, fairness, privacy and others. They will also focus on collaboration between humans and AI, transparency, and robustness.