FIGURE 4.

Displays the accuracy of the model (percentage of correct inferences over 30 trials) under different levels of precision for two parameters (denoted by temperature values for a softmax function controlling the specificity of the A and B matrices for hidden state factor 1; higher values indicate higher precision). As can be seen, the model performs with high accuracy at moderate levels of precision. However, its ability to infer its own emotions becomes very poor if the precision of either matrix becomes highly imprecise. Accuracy here is defined in relation to the response obtained from an agent with infinite precision – and can be taken as a behavioral measure of the quality of belief updating about emotional states. These results illustrate how emotion concepts could be successfully inferred despite variability in lower-level observations (e.g., contexts, arousal levels), as would be expected under constructivist theories of emotion (Barrett, 2017); however, they also demonstrate limits in variability, beyond which self-focused emotion recognition would begin to fail.