Fig. 5.

The generative model used to simulate reading. This graphical model shows the conditional dependencies of the generative model used in subsequent figures, using the same format as Fig. 1. In this model there are two hierarchical levels with three hidden states at the second level and four at the first level (hidden states and outcomes pertaining to categorical decisions and feedback have been omitted for clarity). The hidden states at the higher level correspond to the sentence or narrative – generating sequences of words at the first level – and which word the agent is currently sampling (with six alternative sentences and four words respectively). These (higher level) hidden states combine to specify the word generated at the first level (flee, feed or wait). The hidden states at the first level comprise the current word and which quadrant the agent is looking at. These hidden states combine to generate outcomes in terms of letters or icons that would be sampled if the agent looked at a particular location in the current word. In addition, two further hidden states provide a local feature context by flipping the locations vertically or horizontally. The vertical flip can be thought of in terms of font substitution (upper case versus lowercase), while the horizontal flip means a word is invariant under changes to the order of the letters (c.f., palindromes that read the same backwards as forwards). In this example, flee means that a bird is next to a cat, feed means a bird is next to some seeds and wait means seeds are above (or below) the bird. Notice that there are outcomes at both levels. At the higher level there is a (proprioceptive) outcome signalling the word currently being sampled (e.g., head position), while at the lower level there are two outcome modalities. The first (exteroceptive) outcome corresponds to the observed letter and the second (proprioceptive) outcome reports the letter location (e.g., direction of gaze in a head-centred frame of reference). Similarly, there are policies at both levels. The high-level policy determines which word the agent is currently reading, while the lower level dictates the transitions among the quadrants containing letters.