| The basal ganglia (BG) are implicated in decision making, reinforcement learning, learned conditional gating of orientation movements, and switching among learned behavioral alternatives. Extant computational models embody shared and conflicting hypotheses regarding the division of labor among BG and frontal cortex, and each omits a different subset of the real system's interactions. New computational models are developed and simulated here to offer more complete treatments of local circuits, based on recent neuroanatomical and neurophysiological data. First presented is a model of how firing properties of striatal cholinergic interneurons emerge, under several behavioral paradigms and pathological conditions, from interactions among afferents to these neurons and their intrinsic properties. The model embodies the hypothesis that the internal teaching signal that modulates reinforcement learning at cortico-striatal synapses is a cascade involving dopamine (DA) and acetylcholine (ACh) signals. A second new model, based on dendritic interactions between transmitters of opposing valence in substantia nigra (SN), shows how sustained DA signals that indicate conditional uncertainty can coexist with phasic DA signals that indicate reward prediction errors. This model suggests a novel role, in the genesis of DA responses to probabilistic schedules of reward, for a SN zone with the brain's highest concentration of the neuropeptide substance P A third model is developed to explicate the potential roles of many features heretofore omitted from most BG circuit models, notably up- and down-states of striatal projection neurons, electrotonic communication via gap junctions, and feedback from the globus pallidus to the striatum. This model embodies several novel hypotheses while subsuming the previous two models and key features of other recent models. It explicates complementary functions of midbrain dopaminergic and intralaminar thalamic afferents to striatum during decision sequences, and shows how these and cortical afferents can interact with intrinsic BG circuitry to support adaptive choices, notably by helping to weigh the utility of completing an ongoing plan versus the utility of abandoning it in favor of a promising alternative. These hypotheses and the corresponding computational models integrate a wide range of multi-disciplinary data, and make many empirically testable predictions regarding cognition, learning, performance, and bases for neural responses. |
