Artificial intelligence combined
Artificial intelligence expands into all areas of the daily life, including research. Neural networks learn to solve complex tasks by training them on the basis of enormous amounts of examples. Researchers at the Max Planck Institute for Astrophysics in Garching have now succeeded in combining several networks, each one specializing in a different task, to jointly solve tasks using Bayesian logic in areas none was originally trained on. This enables the recycling of expensively trained networks and is an important step towards universally deductive artificial intelligence.
In astrophysics, artificial intelligence (AI) classifies galaxies, stars, and other objects. AI systems help to control telescopes and analyze their data. They process amounts of data no human would even remotely able to handle. In addition, AIs are increasingly used outside of research in almost all areas, from streaming services to provide users with tailor-made suggestions, from autonomous driving to diagnostic systems in medicine.
The training of an AI is complex and expensive. A sufficiently large and pre-classified data set must be assembled. Based on this, a neural network learns to perform a certain task, e.g. estimating the age of a depicted person. AIs can also be trained to generate realistic examples, exhibiting characteristics of the training data, for example high-quality portraits. During the intensive learning process, networks internalize characteristic features and concepts of the studied faces or other objects. Thus, the networks become representatives of the trained concept. Making the insights of the trained networks available for other tasks is a subject of current AI research.
In the Information Field Theory Group at the Max Planck Institute for Astrophysics, the researchers Jakob Knollmüller and Torsten Enßlin have now succeeded in combining already trained networks in a way to jointly master tasks, none would have been able to on its own. The manner of combining the networks is completely generic and can be used for many different applications without having to train a new network. A so-called generative network is intelligently combined with one or more classifying networks to generate examples that fulfill the required properties. This way elaborate questions can be asked.
For example, an AI for generating faces can be combined with AIs for determining the age and gender of photographed persons. The combined AI then generates a set of possible images of persons that are consistent with incomplete and noisy data of a face (Fig.1 and 2). Since there is usually no certainty on the correct solution for tasks of this kind, this set of images contains the AI's answer. From them, an average image and its uncertainty can be calculated if required (Fig. 3).
To integrate different types of information, the combined AI employs the so-called Bayesian logic, which uses probabilities instead of the binary “true” or “false” of mathematical logic. Bayesian probabilities allow to include uncertain knowledge; in the example here, this would be the information that the person is probably a woman and about 30 years old. Bayesian probabilities optimally support the handling of incomplete and noisy data, such as the rough input image. The various specialized neural networks used to solve a task enter the procedure through probability functions.
The idea of combining AIs with Bayesian logic is not new. However, technical difficulties have until now prevented its realization. The Garching-based researchers were able to overcome this hurdle thanks to a new method originally developed for improved image reconstruction in astronomy. This procedure, called Metric Gaussian Variational Inference (MGVI), allows to perform very large reconstructions, with millions upon millions of unknown quantities, without losing sight of their manifold interdependencies. A first application of MGVI was the three-dimensional reconstruction of the distribution of galactic dust using information field theory (see Highlight: February 2019).
The researchers have now shown how MGVI can be used to combine individual, highly specialized AIs into logically deductive and thus versatile intelligences. In addition, imaging procedures in astronomy, medicine, and other fields can now directly access expert knowledge stored in AIs without having to re-train these on the characteristics of a new measuring instrument. These expert AIs can contribute their knowledge in various specialized applications, such as tumor detection in medical imaging.