MIT to Create 'AI Physicist'

System can generate theories about physical laws in fictional universes
06 November 2018   641

Researchers at the Massachusetts Institute of Technology (MIT) created the AI ​​Physicist system, which is capable of generating theories about physical laws in fictional universes. This will allow AI to extrapolate its knowledge and predict the future.

Artificial intelligence is not yet able to recognize objects or situations, discarding irrelevant details. In other words, it does not know how to focus on a particular object. The reason is that AI at the current level of development is unable to determine what is necessary. For example, if you show it many photos of cats, where they will be in different environments, the system will not be able to identify them because of this difference.

MIT used a different approach. Researchers Tailin Wu and Max Tegmark have programmed four strategies in the machine learning algorithm that scientists use to generate theories about complex observations. They also added a method of small models. These models describe a certain subset of objects, and then a larger “theory of everything” is formed from them.

These are the techniques: divide and conquer (generation of multiple theories, each of which explains part of the overall picture), Occam's razor (using the simplest theory as much as possible without involving third-party entities), unification (combining theories) and lifelong learning (trying to apply theories to solve future problems). AI introduced a series of progressively more complex virtual environments with unusual and strange physical laws. The task of the machine was to predict the behavior of objects in these environments.

In order for the AI ​​to determine how objects will move in two dimensions in these environments, it had to create his own physical theories. AI Physicist, as reported, was able to predict the behavior of the ball in an environment with different physical phenomena in more than 90% of cases, which is much higher than that of traditional machine learning systems.

It is assumed that in the future, AIs will be able to independently set tasks and conduct experiments, even in virtual space. This will allow scientists to better understand complex systems, as well as use artificial intelligence to predict climate change, the economy, and other systems with large amounts of data. As for science, such systems will surely find application in astronomy, physics, chemistry, and so on.

Nvidia to Open SPADE Source Code

SPADE machine learning system creates realistic landscapes based on rough human sketches
15 April 2019   656

NVIDIA has released the source code for the SPADE machine learning system (GauGAN), which allows for the synthesis of realistic landscapes based on rough sketches, as well as training models associated with the project. The system was demonstrated in March at the GTC 2019 conference, but the code was published only yesterday. The developments are open under the non-free license CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0), allowing use only for non-commercial purposes. The code is written in Python using the PyTorch framework.

Sketches are drawn up in the form of a segmented map that determines the placement of exemplary objects on the scene. The nature of the generated objects is set using color labels. For example, a blue fill turns into sky, blue into water, dark green into trees, light green into grass, light brown into stones, dark brown into mountains, gray into snow, a brown line into a road, and a blue line into the river. Additionally, based on the choice of reference images, the overall style of the composition and the time of day are determined. The proposed tool for creating virtual worlds can be useful to a wide range of specialists, from architects and urban planners to game developers and landscape designers.

Objects are synthesized by a generative-adversarial neural network (GAN), which, based on a schematic segmented map, creates realistic images by borrowing parts from a model previously trained on several million photographs. In contrast to the previously developed systems of image synthesis, the proposed method is based on the use of adaptive spatial transformation followed by transformation based on machine learning. Processing a segmented map instead of semantic markup allows you to achieve an exact match of the result and control the style.

To achieve realism, two competing neural networks are used: the generator and the discriminator (Discriminator). The generator generates images based on mixing elements of real photos, and the discriminator identifies possible deviations from real images. As a result, a feedback is formed, on the basis of which the generator begins to assemble more and more qualitative samples, until the discriminator ceases to distinguish them from the real ones.