Google trains artificial intelligence systems to solve problems in a humane way

Google's DeepMind researchers have formed automated learning systems to pbad IQ tests designed to measure a number of thinking skills to demonstrate their ability to think abstract concepts, making the systems more efficient. Artificial intelligence capable of abstract thinking. Like human beings, they can think and develop solutions to problems.

Neural network-based models continue to offer astonishing results in solving automated learning problems, while demonstrating their ability to think that abstract concepts have proved difficult over the course of the day. . .

The researchers published a research article titled "Measuring Inductive Inference in Neural Networks", detailing their ability to measure different abstract thinking abilities of artificial intelligence systems based on IQ used to measure the capabilities of abstraction.

The riddles of the test include a series of random shapes that participants must study to determine the rules that complete this pattern Once the rules of the puzzle have been developed, they must be able to select the following figure in the order .

Deep Mind researchers hope that developing an artificial intelligence capable of going off the beaten track can allow machines to create new solutions to problems that humans have never thought of.

Deep Mind Researchers Used Puzzles Raven Progressive Matrices, These tests were developed primarily by John C. Raven in 1936, where test takers are asked to learn about the subject. ;element. Matrices measured participants' ability to understand the meaning of complex or confusing data.

The researchers developed a software system capable of generating unique matrices to apply this test to artificial intelligence systems, then they formed artificial intelligence systems to solve this test, which allowed to obtain an accuracy rate of 63% in solving IQ style puzzles.

They also tested the ability of systems to recognize new patterns and relationships and to form largely nonverbal structures that facilitate complexity management. "Abstract reasoning is important in areas such as scientific discoveries where we must impose new hypotheses and then use these hypotheses to solve problems. It is important to note that the purpose of this work is not to develop a network of neurons. .

The people who attend the tests can give themselves a boost thanks to the intensive preparation of the test: they learn the type of rules used to control the patterns, which means instead of using an abstract knowledge .

While artificial intelligence systems that use artificial neural networks feed large amounts of data to learn, they can easily learn to capture these patterns without having to use abstract thinking.

The researchers therefore tested a set of standard neural networks on a single property within one matrix but not all of the possible characteristics, and they found a very poor performance where precision did not occur. was only 22% . Who can conclude relationships between different parts of the puzzle recorded the highest accuracy of 63%.

Due to the design of the tests, it was not possible to compare these scores directly with people as artificial intelligence systems were previously trained on how to solve them . They can score more than 80% accuracy while new testers often fail to answer all questions.

This test suggests that existing AI systems will not be able to solve tasks that are not trained to solve them, which means that they need more time and control to achieve it.

In recent weeks it has been revealed that Google's artificial intelligence could now recognize its unique image-based environment, forming systems on visual and cognitive tasks using large sets of images Annotated produced by humans, The Intelligent System, developed as part of Deep Mind, entails viewing any space in a static image called the Generative Query Network, a framework through which systems teach their environment in a meaningful way. only resulting in data obtained from themselves a traveling to a remote location, where the GQN system learns by putting his remarks to understand the world around him. To do this, the GQN system learns to recognize scenes and their geometric properties without putting human marks on the content of these scenes.

The GQN system gives the machine a "human imagination," which allows the algorithm to generate three-dimensional prints of distances that you have never seen in two-dimensional flat images. Deep Mind, CEO of Deep Mind, has announced a breakthrough in artificial intelligence systems with GQN. Dr. Hasabis and his team have been trying to replicate the way the human brain recognizes its environment by examining it. By manually naming the data and slowly inserting it into the artificial intelligence systems, and by forming Deep Mind neural networks, the team presented static images taken from different perspectives on the same stage on artificial intelligence systems. To predict the emergence of something in a new perspective are not included in the photos and the systems quickly learned to imagine a full three-dimensional images of the scene.

Intelligent machines are therefore able to move in the area you imagine, but as they move, the algorithm must constantly make predictions about where the images are seen for the first time.

The researchers published an article explaining their findings: "It was not at all clear that the neural network could ever learn to create images in such a precise and controlled way, but we discovered that the networks deep could learn more about the general perspective. Closures and lighting, without any human engineering, and it was a very spectacular result. "

To create these complete scenes, the system uses two components: the first component manages the representation and creates three-dimensional characters in the static image and transforms it into a complex mathematical form This is known as vector vector .

The second component is called Generative which uses the first component vectors to imagine what is different in this scene – to identify the parts not included in the original images – and so artificial intelligence systems are able to determine spatial relationships within the scene, using data collected from primary images.

Google's sophisticated artificial intelligence systems can also control things in this virtual virtual space by applying their understanding of spatial relationships to the scenario.