Basal ganglia models

Lens adapted to this topic: Leading models and the evidence behind them

The mainstream modeling view treats the basal ganglia as a set of anatomically connected loops whose dynamics help select, reinforce, suppress, and specify actions. Influential frameworks use direct and indirect pathways, disinhibition, dopamine-dependent plasticity, and reinforcement-learning concepts. Reviews also caution that these models are hypotheses: evidence has supported some predictions while conflicting with or requiring revisions to simplified versions.

• Action-selection models, including the GPR framework, explain basal ganglia function in terms of selecting among competing motor programs or behaviors.
• Actor–critic models map dopamine-neuron activity onto temporal-difference prediction errors and striatal plasticity onto reward-guided action learning.
• Contemporary reviews argue that basal ganglia functions extend beyond movement to cognition and affect, requiring models broader than a purely motor account.
• Spiking models test how pathway structure, synaptic plasticity, neuromodulation, and cell-level dynamics could produce behavioral learning.
• The field evaluates models by comparing their predictions with anatomical, physiological, animal, and human evidence rather than treating any one canonical diagram as definitive.

Lens adapted to this topic: Dissenting approaches and unresolved limitations

A serious alternative view holds that canonical box-and-arrow, direct/indirect-pathway, or discrete-action models are too coarse to represent the basal ganglia’s internal integration, territorial organization, overlapping action representations, and context-dependent action specification. This perspective favors multiscale, topological, embodied, or richer spiking models, while recognizing that newer computational proposals remain model-based interpretations rather than settled descriptions of the brain.

• Architectural analyses challenge treating nuclei as homogeneous boxes, emphasizing motor, associative, and limbic territories and integrative properties within the basal ganglia.
• The discrete, non-overlapping action units used in traditional models may be inadequate for explaining how actions are specified, motivating models with richer representations.
• Large-scale spiking-network models incorporate cell subtypes, synaptic dynamics, neuromodulation, and anatomically derived connectivity to move beyond highly simplified pathway diagrams.
• Embodied robot models argue that action selection and behavior switching should be assessed through interaction with environmental demands, not only through internal circuit calculations.
• Alternative computational approaches use detailed pathway models and reward-dependent plasticity to test whether particular circuits are functionally necessary for learning, rather than assuming their roles from anatomy alone.