Dubai: the gate Arab news technical
Researchers at the corporation deb of Maine DeepMind, a subsidiary of Google training machine learning systems to pass IQ tests IQ which is designed to measure a number of thinking skills, in order to prove their ability to think in abstract concepts, thus becoming the systems of artificial intelligence capable of abstract thinking like humans and it can interact and develop solutions to the problems.
Using models based on neural networks in achieving amazing results in solving the problems of machine learning, as proof of their ability to think in abstract concepts has proved to be difficult days, but today has exceeded the artificial intelligence this problem through the training on intelligence tests IQ.
Researchers published a research paper titled “measuring the inference of the Council in neural networks”, explain in detail their attempt to measure the ability of understanding abstract various systems of artificial intelligence based on IQ tests IQ is used to measure the capacity of abstract thinking in humans, which revealed some insights important.
Includes puzzles existing in the test series of random shapes, which participants need to study them to determine the rules that complement this pattern, once the rules are put the puzzle should be able to choose the shape, following strictly in the registry.
Hoped researchers deep mine in that the development of artificial intelligence capable of thinking outside the box can lead to to become machines able to create new solutions to problems that you never think of humans.
Have been used researchers deb Mayne puzzles known as a matrices test. raven cascading Raven Progressive Matrices, the phase of these tests primarily by John C. raven, John C. Raven in 1936, where participants are asked in the test to identify the missing element that completes a pattern, and show a lot of styles in forms of matrices, and the matrices measuring the ability of participants to understand the meaning of complex data or confusing.
Researchers were able to develop a software system designed specifically for this task and is capable of generating matrix of the unique development of this test on artificial intelligence systems, then trained artificial intelligence systems to solve this test, with the result that the artificial intelligence systems achieved a rate of accuracy of up to 63 per cent in solving puzzles test IQ IQ-style puzzles.
As they they were testing the ability of systems to recognize patterns and new relationships forming the structure of the non-verbal to a large extent make it easy to handle complexity.
Said David Barrett, of deb of Maine: “the reasoning of the Council is important in areas such as scientific discoveries where we need to impose new hypotheses, and then use these hypotheses to solve problems, it is important to note that the goal of this work is to develop neural network that can pass IQ tests only”.
Can humans who sit in the tests to give themselves a boost through the intense preparation later, they learn the type of rules used in the control patterns used in the matrix, this means instead of using the abstract idea that they use the knowledge they have learned instead.
While artificial intelligence systems that use neural networks feeder in huge quantities of data to move, can easily be related only to pick up these patterns without the need to use abstract thinking.
So the researchers tested a group of neural networks the standard on one property inside the matrix but not all the characteristics, and found that they perform quite bad performance where the ratio of the precision to 22% only, however the neural network specially designed which can conclude the relationships between different parts of the puzzle recorded the highest accuracy by 63 per cent.
Due to the design of the tests, it was not possible to compare these scores directly with the people because the systems of artificial intelligence have prior training on how to solution, but the researchers found that the participants who have a lot of experience in tests and who could compare them with machinery, can record the rate of accuracy of more than 80 per cent, while it often fails newcomers to the tests in answer to all the questions.
Refers this test to artificial intelligence systems currently existing will not be able to solve the task and not trained to solve, this means that they need more time and observation to get to it.
While in recent weeks it was revealed that the artificial intelligence of Google can now exposed to the surrounding environment based on a single image, through training systems to professional visual and using large data sets of annotated images produced by humans, where the intelligent system that has been developed as part of the deep mine trained himself to visualize any distance in a static image, which dubbed the network the query experimental Generative Query Network, It is the framework through which education systems perceive their surroundings through training only to data which is obtained from themselves during their movement in the place, where the system GQN through the development of his remarks to understand the world around him. And to do so with respect to system GQN identify the scenes and properties of Engineering without the development of any signs of human on the contents of these scenes.
The system gives the GQN machine “silhouette similar to the human being”, that allows the algorithm to generate the impression of three-dimensional spaces that haven’t seen her at all in Flat images two-dimensional.
This was announced Demi photo ass CEO deep mine of the boom achieved by artificial intelligence systems, with the system GQN tried Dr easy and his team replicate the way the brain learns the human to its surrounding environment when you look at it, this is a completely different approach to most projects where the researchers say naming the data manually and enter them slowly to artificial intelligence systems, and to train neural networks, a subsidiary of deb of Maine, the panel view still images taken from different views on the same scene on the systems of artificial intelligence, Using these images enables the algorithm to teach itself how to predict the emergence of something in a new point of view is not included in the photo, and quickly learn the systems perception of three-dimensional images of the entire scene.
Consequently, becoming intelligent machines able to move in the That you imagine, but during the move it must be that the algorithm predictions are constantly about the place of things that are shared in the beginning in the photo and what they look like from the perspective of lasting change.
Researchers published a research paper describes their findings, said researcher Ali Islamic: “it wasn’t at all clear that the neural network can learn to never invent images in a precise manner and controlled like this, however we found that networks deep can you learn about the product general manager of closure and lighting, without any engineering and human, this was the result of a very spectacular”.
To create these scenes the full use of the system component, the first component treats the representation and job codes for the three-dimensional still image and convert it into the form of a mathematical complex, and this is what is known as the vector vector.
The second place is called generative, which uses the first component of vector vectors to guess what’s different in this scene – meaning identify the parts of gas embedded in the original image – building have become the artificial intelligence systems able to ascertain the spatial relationships within the scene, using the data collected from the initial images.
How can systems of artificial intelligence evolving from Google also control the things inside this virtual space imagined through the application of their understanding of the spatial relationships of the scenario.
Link to it from the source: Google artificial intelligence systems to solve problems in a human manner