Brain-based devices: An embodied approach to linking nervous system structure and function to behavior

IN the last several decades great progress has been made in describing how the separate systems of the brain function. Less progress has been made in obtaining a global picture of higher brain functions such as learning and memory. One of the obstacles in this research is that it is currently infeasible to completely enumerate the anatomy or to simultaneously record the activity of all neurons of an animal’s nervous system. Faced with the complexity of real nervous systems our research group has taken a theoretical approach known as synthetic neural modeling By producing simulated nervous systems with great levels of anatomical and physiological detail we can start to address the questions of how structure and dynamics can give rise to higher brain functions. Synthetic models allow us to perform analyses that are currently impossible in real animals, because we have access to the entire anatomy and state of the simulated nervous system. By virtue of the homology between our simulated nervous systems and those of real mammals, we can make predictions for neuroscientists studying behaving animals. However, it is important to realize that the nervous system does not function in isolation. The brain is embodied, and the body is embedded in the environment. The properties, mechanisms, and functions of the nervous system arise only as it interacts with the rest of the body and as the animal engages in a behavior. To that end we engage our synthetic neural models with behavioral tasks and embody them in robotic phenotypes. We call this kind of model a brain-based device (BBD). A BBD can be considered a neurally-controlled cognitive robot. However, it is different from many other similar approaches by virtue both of its construction and its purpose. In contrast to robots that are controlled by what is commonly referred to as an artificial neural network, BBDs have startlingly more complex and realistic neural systems. Typically an artificial neural network may have hundreds of neural units with relatively simple dynamics arranged in a largely feedforward, layered architecture. In contrast, the simulated nervous system of a BBD typically has 10 – 10 simulated neuronal units with much more biologically plausible dynamics, connected with 10 – 10 synapses arranged in a massively re-entrant architecture that mimics the macro-