How many green areas are there near the office? What is the best restaurant in the area? These are everyday questions that require the ability to reason, relate and create connections , a basic aspect of sophisticated thinking that Artificial Intelligence (AI) had not managed to handle. Now, Google’s DeepMind researchers have developed a simple connective reasoning algorithm that has outperformed humans in a complex image comprehension test.
People are generally good at that type of reasoning capable of relating through logic that allows us to connect and compare places, sequences and other realities . But the two main classes of AI – statistical and symbolic – have been slow to develop that ability. Statistical AI (machine learning or machine learning) is very good at pattern recognition, but not at using logic. For its part, symbolic AI is capable of reasoning and establishing relationships using predetermined rules, but it is not good at learning as it goes.
Google’s DeepMind study proposes a new path: an artificial neural network capable of relating concepts . In the same way that neurons are connected to the brain, neural networks are assembled together using small programs that collaboratively find patterns in data . And they can have specialized structures to process images, analyze language or even learn games.
The relationship network is designed to compare pairs of objects in different scenarios individually . According to Timothy Lillicrap, a computer scientist at DeepMind in London and a co-author of the research, “we are explicitly forcing the network to find the relationships between objects.”
This expert and his team challenged the network to carry out various tasks, such as answering questions about the relationships between objects – cubes, balls, cylinders – that appeared in the same image.
For example: There is an object in front of a blue thing. Is it the same shape as that other red thing to the right of the gray metallic ball? To solve this problem, the network capable of relating worked in combination with other neutral networks, such as one capable of recognizing objects in the image and another capable of interpreting questions . In the total of tests and questions posed, other machine learning machines got between 42 and 77% correct, the humans who participated reached 92% and the new relational network got 96% of the questions correct .
This new AI network also shone in language exercises, particularly inference exercises. For example, they would ask, “Lily is a swan; Lily is white; Greg is a swan; what color is Greg? (White).” In this type of exercise, their level of success was 98%, compared to 45% in other AI programs .
It is a great advance, but according to experts, to achieve the flexibility of human intelligence, the network must learn by example to compare not only pairs of things, but trios, double pairs and other combinations.
