Basal ganglia circuit model: rostral vs caudal — population responses to learned object value
Signals flow Str → GPe → STN → SNr → SC, as drawn. The model includes GPe→STN and rostral STN→GPe connections, and separate dopamine neuron groups projecting to the rostral and caudal striatum (Kim et al., 2015, Cell). Arrow colors: inhibitory (blue), excitatory (red), dopaminergic (green).
Developed by Hyoung F. Kim & team members · cocila, Seoul National University (SNU) · Released October 14, 2026 · If you use or adapt this model, please cite the source accurately (see "About this model").
About this model · how it was built
What it is
A firing-rate model of two parallel basal ganglia–superior colliculus loops that guide saccades to objects by their learned value. The rostral circuit (caudate head side) carries flexible value learned in a reward-reversal task; the caudal circuit (caudate tail side) carries stable value and is tested in passive viewing without reward. Both circuits converge on the SC, whose activity is read out as saccade choice and reaction time.
How it was built
Architecture. Striatum (value-coding Str/pos and Str/neg populations) → GPe → STN → SNr → SC, with a cortical (hyperdirect) input to STN and a direct GPe→SNr projection. Connections follow known anatomy for each region (e.g. caudal: positive-value CDt → SNr, negative-value CDt subset → GPe, no STN→GPe). Each nucleus is a leaky rate unit with a softplus transfer function anchored at its recorded baseline rate, with millisecond-scale delays.
Learning. Object-specific delta-rule learning of the striatal weights and of the cortex→STN synaptic efficacy. Both circuits use the same rule and differ only in learning rate (rostral 0.12, caudal 0.0011); unrewarded passive viewing acts as extinction. Cortical firing itself is value-neutral; learned value is stored in synaptic efficacy.
Fitting. Circuit parameters were fitted, within physiological bounds, to recorded population PSTHs (rostral: caudate head, GPe, STN and SNr in the reversal task; caudal: caudate tail, ventral GPe, STN and dorsolateral SNr in passive viewing), together with pathway-block results (rostral: GPe→STN block ineffective, cortex→STN block effective; caudal: both effective).
Output and behavior. SNr→SC weights were fitted to recorded SC responses. Behavior is produced by a leaky accumulator race on the SC signal, calibrated only to intact saccade reaction times, reversal choice accuracy and passive-viewing look rates. Behavior under lesions is a model prediction, not a fit.
Virtual lesions. Any nucleus or pathway can be silenced partially (10–100 %) in either circuit; effects are measured as the change in SNr Good–Bad value discrimination and in behavior.
How to use
Choose the task (Reversal / Passive viewing) and the view (Rostral / Caudal / Rostral + Caudal / Behavior), then press Play trials.
Select lesions in the lesion panel; the readout cards and PSTHs update with the lesioned circuit.
Longer explanations under each panel are folded; click "Show details" to open them.
Credits and use
Developed by Hyoung F. Kim & team members at cocila, Seoul National University (SNU). First public release: October 14, 2026.
If you use, adapt, or show this model (including screenshots, figures or videos), you must state the source clearly and accurately. Please cite:
Recorded PSTHs were digitized from population plots (about ±2–3 Hz). The model is a population-level rate model; spiking, oscillations and detailed visual pathways are not modeled, and parameter values outside the fitted conditions should be read as predictions.
Note: each Rostral/Caudal row has Glu→STN (block of the glutamatergic cortical/hyperdirect input) and GPe→STN (block of the pallidal inhibitory input) buttons. A pathway block does not lesion the STN itself; it only sets that input to zero, and it acts on the full dynamics of each circuit independently. Each button toggles independently, and rostral and caudal lesion states are stored separately. STN/GPe lesions act on the same circuit regardless of task; reversal–PV differences arise from each circuit's learned state and dynamics, with no task-specific gate. An STN lesion removes only the STN's own downstream contribution and does not weaken Str→SNr or GPe→SNr signals. A DA lesion blocks learning at that site (η→0). Lesion effects are evaluated as the reduction of SNr Good–Bad value discrimination, not at the SC. No lesion-specific behavioral penalty is used.
SNr value discrimination · behavior readout
Shown after a simulation is run.
Method / caveats
No fitted choice/RT equation and no lesion-specific behavioral penalty. Behavior comes from the SC saccade race: RT = single-object saccade latency (Good, Bad); choice = two-object competition between the Good and Bad SC channels. One parameter set for every tab, task, region, intact and lesion condition; the only task difference is fixation-related inhibition (reversal: fixation held 0–400 ms, go at 400 ms, then inhibition decays; PV free viewing: constant gaze engagement). In PV the monkey may not look at an object at all, so PV behavior = look probability and gaze bias (look Good − look Bad) plus RT of executed saccades. Caveat: calibrated on intact RT targets (reversal saccade Good≈450/Bad≈550 ms after object onset (400-ms fixation hold), PV mean≈150 ms, PV look Good≈60%/Bad≈40%) and reversal choice≈85% only; gaze bias and all lesion behavior are predicted, not fitted.
Average PSTH (n = 25 trials, mean with trial-to-trial noise)
Mean firing rate of each node for the good (red) and bad (blue) object. x axis −200 to 600 ms, y axis spikes/s. The gray bar marks stimulus presentation (0–400 ms).
Each panel shows the Good (red) and Bad (blue) response of one value-coding subpopulation. Ratios are empirical neuron-count constraints; the aggregate PSTH is computed only as the neuron-count-weighted mean of the forward-simulated Pos/Neg populations.
rGPe Positive (+) · 6/7
rGPe Negative (−) · 1/7
rSTN Positive (+) · 1/2
rSTN Negative (−) · 1/2
rSNr Positive (+) · 1/10
rSNr Negative (−) · 9/10
Weighted-mean check: rGPe = (6·GPe+ + 1·GPe−)/7; rSTN = (STN+ + STN−)/2; rSNr = (SNr+ + 9·SNr−)/10. This panel is shown for the rostral circuit only and does not affect the caudal circuit.
Note: value selectivity first appears in tail Str, then rapidly in cGPe/cSNr, followed by cSTN and finally SC. The exact onset values are approximate visual estimates from the supplied population PSTHs, so the model uses conservative relative pathway delays rather than claiming single-neuron conduction times.
Key references:
Kim & Hikosaka (2013) Neuron 79:1001–1010;
Griggs, Amita, Gopal & Hikosaka (2018) Frontiers in Neuroscience;
Kim (2025) Brain 148:4198–4209.
Lee et al. (2026) bioRxiv.
Passive-viewing circuit
Passive-viewing simulation uses the same core circuit with caudal stable-value dynamics. STN receives no direct input from Str in either mode. In the Rostral circuit, GPe->STN inhibition is reciprocated by a positive STN->GPe excitatory arm; in the Caudal stable-value circuit used here, only GPe->STN is included and no cSTN->cvGPe arm is modeled. Lesions selected above are applied to this circuit.
Passive-viewing task: 8 sessions
Structure: learning (100 rewarded trials) → PV test (a 25-trial viewing session with no reward at all) → learning → ...
The PV test itself consists of many trials, and the absence of reward on all of them acts as extinction under the delta rule.
Each session below therefore shows the state after the unrewarded PV session, not the moment learning ended — the fast-learning (rostral) circuit is strongly extinguished, the slow-learning (caudal) circuit hardly at all. Important: there is no task gate that switches rostral or caudal on or off. Both circuits receive the same R = 0 update; the difference comes from η alone.
Each cell is the mean firing rate for the good (red) and bad (blue) object. x axis −200 to 600 ms, y axis spikes/s; the gray bar marks stimulus presentation (0–400 ms). Switch between the Rostral/Caudal tabs above to compare.
SC saccade race (behavior)
Each object has its own SC channel (fused rSNr + cSNr input), and a saccade is made when the accumulator of the saccade-generating stage reaches threshold. In a 1-object trial the time to threshold is the RT; in a 2-object trial the Good and Bad channels compete through mutual inhibition and the first to reach threshold is chosen. Both tasks use the same parameters; the only difference is the physical task structure, i.e. fixation-related inhibition (reversal: fixation must be held for 0–400 ms, so inhibition is strong (Fhold) and decays after the go signal at 400 ms; PV free viewing: gaze stays engaged, so inhibition is constant). In PV the monkey may not look at an object at all (no saccade within 0–400 ms), so PV behavior is evaluated as look probability and gaze bias (Good − Bad). Traces below are Good (red) and Bad (blue) 1-object trials; solid = intact, dashed = selected lesion. Reversal lesions are applied to the rostral circuit, PV lesions to the caudal circuit.
Reversal task
Passive-viewing (habitual) task
Combined rostral + caudal circuits
Rostral and caudal basal-ganglia circuits operate simultaneously. Learned value remains in each Str pathway; downstream nuclei receive only neural activity. The two pathways remain separate through SNr and converge only at the action-specific SC population.
rStr → rGPe/rSTN → rSNr ↓ SCA, SCB
Fast/flexible pathway (η=0.12). Reciprocal rGPe↔rSTN loop is retained; Str has no recurrent dynamics in the executed model.
cStr → cGPe/cSTN → cSNr ↓ SCA, SCB
Slow/stable pathway (η=0.0011). cSTN→cGPe remains omitted as in the current caudal model constraint.
IA/B is the non-value action/cortical drive. rSNr and cSNr are the simultaneously generated outputs of the rostral and caudal circuits. wr and wc are convergence strengths and can be changed below. They are not automatically switched by task.
Each SC trace represents an action/object-selective SC population, not a claim that every rSNr and cSNr neuron converges onto one SC neuron. Both conditions receive the same value-neutral visual burst. Only baseline-relative SNr deviations (ΔSNr) contribute value-dependent inhibition: SNr below baseline disinhibits SC, whereas SNr above baseline suppresses SC. Learned value is stored in Str opponent weights and in object-specific Ctx/Glu→STN synaptic efficacy; cortical firing itself remains value-neutral.
Rostral SNr, caudal SNr, and fused SC output
Press "Run combined simulation" to compute both circuits together.
rSNr PSTH
cSNr PSTH
Fused SC PSTH — good vs bad overlay
SC PSTH — Good object
SC PSTH — Bad object
Current rostral causal subpopulation model: rGPe+/− and rSTN+/− are generated forward from learned Str traces and fixed-sign synaptic weights/time constants. No mirror, inverse solution from the aggregate PSTH, or post-400-ms forced convergence is used. Aggregate PSTHs are neuron-count weighted means of the independently generated populations.
Time convention: object onset = 0 ms. All PSTHs are displayed from −200 to +600 ms. The stimulus is present only from 0 to 400 ms (gray bar above each PSTH); 400–600 ms is post-stimulus activity.
The shading of each box shows that node's instantaneous firing rate (good object) during playback. The graphs are average PSTHs over 25 simulated trials with noise, showing how good (red) and bad (blue) responses diverge over time; they are the final-session responses after 8 reversal sessions (25 trials each). Circuit structure: the striatum is split by value coding into Str/pos (Good > Bad) and Str/neg (Bad > Good) populations; this is not the D1/D2 (direct/indirect) distinction. Each population may have a direct component to SNr and an indirect component to GPe (strengths are region-specific parameters), and the rostral and caudal circuits use the same equations. The STN receives no direct striatal input, only GPe (inhibitory) and a cortical/salience input (excitatory; a common cortical signal at object onset). A direct GPe→SNr inhibitory connection is included; there is no SNr→GPe feedback. The caudal GPe has no independent value-neutral object-onset drive; its early response comes through Str/neg→GPe. Pathway-specific conduction/synaptic delays follow the order of value-discrimination onsets in the recorded PSTHs. The common visual burst of the SC is an effective value-neutral input that compresses the detailed visual pathways; these terms do not read learned value. Dopamine neurons projecting to the rostral and the caudal striatum are anatomically separate groups (Kim, Ghazizadeh & Hikosaka, 2015, Cell); in the model both use the same learning rule (delta rule) and differ only in learning rate (rostral η = 0.12, caudal η = 0.0011). Stable value in the caudal circuit emerges from slow integration; the hypothesis that caudal DA signals stable value itself (Kim et al., 2015) is not implemented separately.
Rostral lesion / pathway-block validation table
All conditions use the same intact parameters with a fixed decoder (intact target accuracy ≈85%). Value discrimination in rSTN comes mainly from the learned Glu→rSTN input; rGPe→rSTN is secondary. This panel refits nothing: each condition removes one nucleus or projection and shows the resulting GPe/STN/SNr Good–Bad discrimination and behavior.
After the page loads, press "Run validation".
η-only ablation
The rostral and caudal circuits share one set of equations; they differ only in the parameter table (REGION). This table crosses circuit parameters with the learning rate η to test where the task dissociation comes from.
Press the button to run.
Rostral population-routing explorer
Fixed principles: learned value is stored in the opponent striatal weights and in the object-specific Ctx/Glu→rSTN synaptic efficacy. Cortical firing itself is a value-neutral object signal, identical for Good and Bad, but the learned efficacy makes the effective drive to rSTN value-dependent. The +/− value-coding populations of rGPe may branch to rSTN and rSNr with different probabilities. Every existing synaptic weight is strictly >0, and there is no Str→STN shortcut. This population-ratio panel is for the rostral circuit only and does not apply to the caudal circuit (whose cSTN is split into GPe- and Glu-recipient populations).
Note: Pos/Neg PSTHs use the same empirical 6:1, 1:1, 1:9 composition and are generated from each nucleus' own causal Good/Bad activity. No downstream node reads Str value directly and no lesion/pathway-specific analysisGain is used. Whole-nucleus lesion forces both subpopulations to 0 Hz.
SC (superior colliculus) output: the SC receives a constant excitatory tonic input and a visual input, and the absolute inhibitory firing rate from SNr: ISC = Itonic + Ivisual − wSNr→SCrSNr. Itonic is set so that the prestimulus SC rate matches the recordings.