Artificial Intelligence is taught abstract thinking

DeepMind is a British company that teaches artificial intelligence to find connections between objects, determine their differences and common features.

Any person uses this type of inference when choosing the best products or building an aggregate picture of the event from a few facts. One of the most powerful tools for understanding the world is the ability to transfer abstract relationships, for example, in the process of analyzing a situation where one object is located on the right or is smaller compared to another one.

But things that person perceives as basics, can be an extremely difficult for machines. It's one thing to do tasks based on image recognition, and it’s totally another to transfer knowledge when analyzing text. If computers manage do this, they will become even closer to the general intelligence.

The company has developed a neural network that specializes in similar type of abstract thinking. It can be connected to another network in order to strengthen the power of reasoning. Scientists have trained artificial intelligence using images of different colors and sizes.

The task of the computer is to analyze the objects and find common features. Other tasks include answers to the questions like: “What is the size of a ball located to the right of a yellow square, which is to the right of a small cylinder?” Currently, the machine answers correctly in 95.5% of cases. Sometimes even people cope worse with the task.

Adam Santoro, the head of the research, notes that it is too early to talk about the application of technology in practice. Most likely, it can be implemented as an application that automatically describes a single image or video. It will be useful and necessary for those who have vision problems.