Brain circuit implementation: High-precision computation from low-precision components

Introduction Attempts to understand, let alone augment or supplant, the operation of brain circuitry rely not only on our knowledge of isolated neuron, circuit, and brain slice behavior but also on the impressive computations achieved by assemblies of these components. The variability of the constituent elements of these circuits suggests that they are arranged to operate in ways not obvious from standard engineering points of view. The quite nonstandard (and in many cases, apparently substandard) unit components of these circuit designs (synapses and cells that are probabilistic, relatively low precision, very sparsely connected, and orders of magnitude slower than the typical elements of most engineering devices) provide constraints on their possible contributions to the overall computation of neural circuits. Since the resultant brain circuits outperform extant engineering devices in many realms of crucial application ranging from recognition of complex visual or auditory signals to motoric traversal of complex terrain, our ability to imitate them can lead to two related but distinct classes of scientific advance: new and unanticipated types of hardware devices based on the unveiled engineering principles, and enabling technologies for the integration of extrinsic devices with intrinsic brain circuitry. This latter capability implies two directions of improved communication: i) a heightened ability to “listen to” and interpret brain activity, and ii) a burgeoning faculty for “talking back” to the brain, which may ameliorate impaired brain function or enhance normal function.