Algorithmic intelligence as an emergent phenomenon

Machine learning deals with processes that learn from examples to perform tasks that cannot be captured in simple rules, such as pattern recognition or autonomous driving. This has a great impact on our working world (keyword "automation") and also on society, e.g. when social media learn our habits. The field has made very great progress, most recently in particular through so-called "deep" artificial neural networks, which often solve pattern recognition tasks in natural data (e.g., photo or speech recognition) so well that they come remarkably close to human capabilities. Interestingly, it is currently poorly understood why deep networks work so well and in so many different applications. The background is that learning from examples is basically only possible if the "learned" patterns come from a relatively small set of possible mathematical patterns. Why natural data all seem to be so similar, and how the networks implicitly encode this pattern and exploit this prior knowledge, is an open scientific question. In the funded project, this question will be addressed from an interdisciplinary perspective: Researchers from physics, biology and computer science are working together to better understand what basic statistical patterns are hidden in natural processes and how modern machine learning methods can learn these patterns. This should lay the foundations for developing even better machine learning methods, on the one hand, and for perhaps better understanding the mysterious phenomenon of intelligence as such, on the other.