P. E. Latham, B. J. Richmond, M. J. O'Donovan, V. Dunlap, and P. G. Nelson
National Institute of Health, NIMH, NINDS, NICHD.
Computations performed by networks of neurons transform one set of spike trains on a group of incoming fibers to a different set on a group of outgoing fibers. From a dynamical systems point of view, a computation corresponds to the evolution of network activity in response to incoming spike trains. One difficulty in understanding computations performed by real neurons is that the evolution of activity in response to input is superposed on ongoing, intrinsic dynamics. We are studying the nature of the intrinsic dynamics, the goal being to understand its effects on input-output transformations. We use isolated neuronal networks (no input), and ask questions such as: What firing patterns occur in isolated networks? Is there a regime characterized by constant firing rate? How are activity patterns related to single cell properties and connectivity?
For our experiments we use dissociated spinal cord neurons. In these cultures no learning has taken place, so connectivity is specified probabilistically. To describe this system, we use a population model similar to that of Wilson and Cowan (1972), but with adaptation included. We make two predictions. First, networks that exhibit low firing rates (1 Hz) must contain endogenously active cells, i.e., cells that fire without any input. Second, network dynamics is characterized by transitions between high and low average firing rates.
Patch clamp recordings from single cells in the presence of a broad array of neurotransmitter antagonists indicate that approximately 10-30% of the neurons are endogenously active. In addition, calcium imaging and two neuron recordings show transitions between high and low firing rate states. Our current hypothesis is that the two states (at high and low firing rates) bifurcate into multiple states as the network learns.