Ph.D. or Post Doc
in Neuroscience and Robotics ??

• Distributed Graphical Cognitive Systems
  A network of individual computational units, each with limited computing power, memory, and communication, which all together solve complex global problems. Such networks might be hand-designed or autonomously constitute their topology and possibly learn or adapt based on sensory input. But all such systems separate their tasks into “localized sub-problems”, where each single actor (a unit in the network) contributes to a global solution, without understanding the complete problem. This style of information processing is very different from today’s computing algorithms, but might reflect distributed information processing in brains.
  
  
• Neural-Inspired Hardware for Information Processing
  Believe Propagation in Factor Graphs (FG) is a standard technique to represent complex relations between variables as computationally simple distributed systems. Such variables can represent e.g. sensory readings or desired motor outputs of robotic systems. Computations in FGs are typically performed on sequential digital hardware (standard PC), which often results in long settling times of FGs with loops. Here, we are envisioning massively parallel hardware systems to compute equilibriums of loopy FGs quickly (such as FPGAs or microcontrollers) or even instantaneously (eg. in analog hardware). This research might provide new insight in high speed motor control problems in humans and robots (eg. robotic arm visual servoing).
  
  
• Event-Based Vision Systems for Robots
  Vision is probably our most complex sensory system: around 1/3rd of the human cortex seems to be devoted to visual information processing. Robotic systems also often include cameras, and typically also need strong computational power to process vision data. Biology, however, uses a trick to simplify processing of visual information: the eye already compresses information at the source. Instead of continuously communicating full images to the brain, it only encodes visual changes as streams of spike-events. We have hardware sensors that work just the same: in contrast to traditional cameras, these sensors only transmit updates - or changes - within their field-of-view. We are exploring their application in several scenarios, which include object tracking, distance sensing, obstacle avoidance, and ego motion stabilization of airborne helicopters.