Parameter selection and state dominance in hidden Markov models of neuronal activity

Hidden Markov model (HMM) analysis of single-neuron activity supports the concept that neural networks in the brain may operate by switching between a relatively small number of stable configurations of activity (attractors) in the processing of specific tasks. Here, we consider two problems: 1) how to estimate the true number of HMM states in the source (which is unknown) from the observed activity; and 2) how to decide whether the states are likely to be associated with a simple distribution such as the Poisson or a more complex one (Gaussian). We also show how state dominance (the observation that at any time one state is much more probable than all the others) depends on the source parameters. To do this, we deal with artificial data generated with known Markov statistics and resembling real neuronal activity