TU Vienna revolutionizes AI: New method makes uncertainty measurable!

TU Vienna revolutionizes AI: New method makes uncertainty measurable!

What's new from research at the Vienna University of Technology? A dedicated research team has developed an innovative mathematical method to calculate uncertainty in neural networks. This method could revolutionize the way we make decisions based on artificial intelligence. In the world of AI, where transparency is often lacking, this is a real bright spot.

The challenge with classic neural networks is obvious: they often deliver the same results with identical inputs, but small changes can produce unexpected results. To date, these systems have not provided any information about the security of their decisions. According to the researchers at the TU Vienna, this is exactly where it is time for a rethink. With the new method, which is based on a geometric approach, the input space is broken down into small sub-areas. The possible outputs can be calculated exactly for each of these areas, which enables the error probabilities to be easily quantified.

Mathematical and technical sophistication

The special activation functions used in ReLU networks are particularly exciting. These are described by two approximation networks, which simplifies the calculation operations enormously. In tests on datasets such as “Iris,” “Wine,” “Diabetes,” and “Banana,” the researchers showed that their method provides more precise mathematical bounds on network outputs than previous approaches were able to. It turned out that the method works particularly efficiently in low-dimensional input dimensions and does not lose accuracy even in more complex challenges - although at the expense of higher computational effort.

A noticeable advantage of this method is its focus on smaller, controllable neural networks. However, more complex AI systems, such as ChatGPT, cannot be analyzed. This insight opens up new perspectives, particularly in safety-critical areas such as medical technology, robotics, the development of autonomous vehicles and financial risk assessment. The overall goal of the developers is to improve decision making by better understanding the probabilities of possible outcomes.

Practical Applications of Research

The potential uses of this new method are diverse and could have far-reaching effects on many areas of life. This technology could soon find application, particularly in fields where safety and accuracy are top priorities. Whether in healthcare or the financial industry – the ability to calculate uncertainties and be able to better justify decisions is very important.

The research work at the Vienna University of Technology not only opens up new possibilities for artificial intelligence, but also provides an important impetus for further development in this exciting and promising area. Researchers and companies are now called upon to adopt this new methodology and think ahead in their own developments. The future of AI could thus become clearer and more predictable.

For more information about the new method and its impact on research and industry, read the detailed reports at Ingenieur.de and TU Vienna.