Core and shadow model. The upper part of the architecture is the core model shown and explained in Figure 3. The states and in the shadow model, displayed in the shaded box, share emission parameters with the corresponding states E
in the core model. This is done such that states in the shadow model do not contribute to estimation of the shared parameters. The labelling of states in the shadow model differs from those of the core model in that the states are now labelled as non-expressed and is labelled as expressed. This allows the Baum-Welch algorithm to use the correspondence between annotation and score in proportion to how likely this is, given the information obtained in the initial training. and states impose a minimum on the number of consecutive probe scores that can be captured by the shadow architecture. This means the model considers at least six consecutive probe scores at a time when evaluating the evidence in the training set. Prior to training the parameters are set so that the probability is distributed uniformly among transitions and among emissions for each state. The exceptions are the transitions shown as dashed lines. These are primed with probability 0.001.