Internal Model Reproduces Anticipatory Neural Activity

Many theorists emphasize the role of an “internal model of the world” in directing intelligent behavior. Internal models predict the evolution of the environment by imitating its causal flow. They compute prediction signals and use these prediction signals to form novel associative chains. Formation of such predictive chains may contribute to reasoning and planning. Animals seem to learn and use internal models; they learn to anticipate predictable events, and their behavior in latent learning experiments reflects formation of novel associative chains. Despite such behavioral evidence, the neural basis of an internal model is still under debate. In order to investigate possible neural correlates of an internal model, a previous model of animal learning was extended to an internal model approach. As believed to occur in animals, the proposed neural network model computes predictions and uses these predictions to form novel associative chains in a latent learning experiment. Simulated signals resembled anticipatory neural activity which has been reported in cortex, striatum, and midbrain dopamine neurons. Simulated prediction signals were comparable to tonic anticipatory activities in cortex and striatum. Furthermore, simulated reward prediction error signals were comparable to phasic activities of midbrain dopamine neurons. These findings suggest that tonic anticipatory activities can reflect prediction signals and that phasic anticipatory activities can reflect prediction error signals. Furthermore, comparison of the model architecture with biological neural networks suggests that chains of neurons with anticipatory activity underlie formation of novel associative chains. In conclusion, anticipatory activity seems to reflect the processing of an internal model.