On a spike train probability model with interacting neural units