Markov transitions between attractor states in a recurrent neural network