Looking at small animals with tiny brains, we get the impression that they act effortlessly in the world, foraging for food and returning home safely. In contrast, today's carefully hand-designed computers and robots with all available sensors and processing power are hardly able to successfully perform such simple behaviors. The world is too complex and too ambiguous to get interpreted reliably with contemporary algorithms. So in which fundamental principles does information processing in brains differ from information processing performed by current computing algorithms?

Probably the most fundamental difference is already established by the design of the elementary unit that performs computation: today’s engineered systems typically rely on relatively few but powerful and cautiously hand-designed processing cores (CPUs) – even high-end machines typically have no more than four CPUs in a system. Brains, in contrast, are composed of a large number of relatively simple processing units (neurons) – ranging in count from a few hundred in the simplest worms up to several 10¹¹ neurons in a mammalian brain. Each such neuron operates with relatively low speed, but all of them work in parallel, forming a large, self-grown, recurrently interconnected network of “computing machines”, each contributing to the overall task. No neuron – and no group of neurons – has access to global information, as CPUs do in our computers.

This difference in computing hardware imposes severe constraints for computing algorithms, that today to a large extent are completely unaddressed. How can a distributed system with only local knowledge perform globally consistent actions? How does such a system build itself – starting from a nucleus – with only local knowledge and no global supervisor? Why is such a large network of neurons relatively insensitive to changes in the connectivity pattern and to defective computing units? How does such a deep network learn?

In my research I am addressing such questions by applying neuronal-style algorithmic primitives to artificial engineered systems that interact intelligently with the real world – thereby working towards understanding how brains perform computation, and ultimately gaining insight in why such systems outperform contemporary algorithms. This page shows past and current research projects, highlighting how I apply principles from neuronal information processing to engineering problems.