Caudate encodes multiple computations for perceptual decisions

Abstract

Perceptual decision making is a complex process that requires multiple computations, including the accumulation of sensory evidence and an ongoing evaluation of the accumulation process to use for prediction and adjustment. Implementing these computations likely involves interactions among many brain regions. For perceptual decisions linked to oculomotor actions, neural correlates of sensory evidence accumulation have been identified in several cortical areas, including the frontal eye field and lateral intraparietal area, and one of their direct, subcortical targets, the superior colliculus. These structures are also connected indirectly, via the basal ganglia. The basal ganglia pathway has been theorized to contribute to perceptual decision making, but the nature of this contribution has yet to be examined directly. Here we show that in monkeys performing a reaction-time visual motion direction-discrimination task, neurons in a primary input structure of the basal ganglia, the caudate nucleus, encode three aspects of decision making: evidence accumulation, evaluation, and choice biases. These results indicate that the basal ganglia pathway can provide important signals to influence and assess perceptual decisions that guide oculomotor behavior.

Figure 1.

Behavioral task and performance. A, The monkey decides the direction of random-dot motion and then responds immediately by making a saccade to one of two choice targets. Saccades to the target in the direction of coherent motion (assigned randomly for 0% coherence) are followed by juice reward. B, C, Example psychometric (top) and chronometric (bottom) functions from a single session for monkey C (B) and monkey F (C). Psychometric functions are plotted as the fraction of trials in which the monkey chose the T1 target as a function of signed coherence, where positive coherence indicates motion toward T1. Chronometric functions are plotted as the mean RT, measured as the time between motion stimulus onset and saccade onset, for saccades toward each target, separated by correct (filled symbols) and error (open symbols) trials. Solid curves are simultaneous fits of both functions to a drift-diffusion model with asymmetric bounds.

Figure 2.

A, B, MR localization of recording sites for neurons included in the analysis. MR images of the coronal plane for monkey C (A) and monkey F (B), with blue lines illustrating estimated electrode tracks and red circles illustrating estimated locations of recording sites for neurons included in the analysis. The anterior–posterior positions of the images relative to the anterior commissure (AC) are indicated by the numbers on the top left (in millimeters, positive: anterior; negative: posterior). The numbers at the top right of each panel indicate the number of cells recorded along the tracks shown. Scale bar, 5 mm. Gray outlines indicate estimated boundary between gray and white matter (see Materials and Methods). CD, Caudate; IC, internal capsule; LV, lateral ventricle; PU, putamen.

Figure 3.

Summary of modulation by choice (A) and coherence (B) across epochs in the population of 129 neurons. Each row represents a single neuron across epochs. Activity in different task epochs was tested separately. Neurons are sorted separately in A and B, as described below. A, Colors represent IN target location of cells showing choice-dependent activity (see color scale to the right). For each neuron, only two motion directions were used and saccade targets corresponded to the motion directions. IN target refers to the preferred target location of the two alternatives. White entries indicate no choice dependence (Wilcoxon rank-sum test, p ≥ 0.05). Neurons are sorted by direction preference in successive epochs. Black circles in the leftmost column indicate cells with bias activity. B, Colors represent the slope of coherence modulation based on a linear regression (see color scale to the right). Positive/negative values imply increasing/decreasing neural responses as a function of coherence. White indicates no significant coherence modulation (F test, p ≥ 0.05). For each epoch, the left and right vertical columns represent IN and OUT choices, respectively. Neurons are sorted by slope values of successive columns in descending order. Black circles in the Stim epoch indicate cells with bias activity during the Pre epoch. Blue and red circles at the right identify evidence accumulation and evaluation activity for each epoch, respectively.

Figure 4.

Example neuron with evidence accumulation activity. A–D, Spike raster plots for correct trials with IN (see text; A, C) and OUT (B, D) choices at 6.4% (A, B) and 25.6% (C, D) coherence. For each panel, the left plot is aligned on motion stimulus onset, and the right plot is aligned on saccade onset. Trials are sorted by RT. Red markers indicate saccade onset. Note that the plots were truncated at 1.4 s after stimulus onset, and saccades occurred after this time in the top few trials in A and B. E, F, Spike-density functions (Gaussian kernel, σ = 20 ms) aligned on motion stimulus (E) and saccade (F) onset. Solid lines, IN trials; dashed lines, OUT trials. Colors correspond to different coherence levels, as indicated in F. Only correct trials were included. In E, spikes after 100 ms before saccade onset were excluded, and the spike-density functions were truncated at the median RT for each coherence level.

Figure 5.

Population summary of neurons with evidence accumulation activity. A, Population average of activity aligned on stimulus onset for correct trials (n = 47). Data for each choice (solid lines, IN trials; dashed lines, OUT trials) and coherence level (colors; see inset in B) were truncated at the median value of the mean RT measured under those conditions for all cells. Activity after 100 ms preceding saccade onset was excluded. B, Population average of activity aligned on saccade onset, plotted as in A. Activity within 200 ms of stimulus onset was excluded. C, D, Population average (mean ± SEM) of normalized firing rate as a function of coherence in 100 ms windows before the median RT (arrow in A, which indicates only an approximate time because of session-by-session variability in mean RT) and before saccade onset (arrow in B), respectively. For each neuron, the firing rate in an epoch was normalized by the average rate across all trial conditions in the same epoch. Filled circles, IN trials; open circles, OUT trials. Lines are linear fits with slopes that were significantly different from zero (p < 0.05). E, F, Histograms of the slope of the rate of change in activity during motion viewing for IN (E) and OUT (F) trials. Solid bars, F test, H₀: slope = 0, p < 0.05.

Figure 6.

Comparison of evidence accumulation activity for error and correct trials with weak motion strength. A, C, Population average of activity from 32 neurons with positive slope_IN for correct (black) and error (gray) trials. Solid lines, IN choices; dashed lines, OUT choices. Data were truncated at the median value of the mean RT measured under each condition for each neuron and standardized by subtracting the average firing rate 200–600 ms after stimulus onset for all trials for each neuron. B, D, Scatter plot of the ROC indices for 3.2% (B) and 6.4% (D) coherence trials. The ROC index was computed comparing the average firing rate 200–600 ms after stimulus onset, excluding activity within 100 ms of saccade onset, for IN versus OUT choices. Values >0.5 indicate greater responses for IN choices. Arrows indicate median values for correct (black) and error (gray) trials. Note that there are fewer data points in D because of fewer sessions with sufficient number of error trials at 6.4% coherence.

Figure 7.

Evidence accumulation activity in the Stim epoch did not converge at a decision bound. A, Regression slope values (thick black lines, right axes) of population average spike density function (color lines, see legends in Fig. 5B, left axes, σ = 20 ms) aligned on stimulus (left panel) and saccade (right panel) onset, with coherence as the regressor for each 1 ms time bin. Colors indicate coherence levels. Red bars on top indicate time bins in which t test of H₀: slope = 0 returned p < 0.05. B, Regression slope values of population average activity aligned on saccade onset, with RT as the regressor. Colors indicate the starting value of each RT bin (bin size 50 ms). The same conventions are used as in A. C, Percentage of neurons with a significant slope from the regression with coherence in the given time bin. Red, Positive slopes; blue, negative slopes; gray areas indicate 5% (chance level for p = 0.05). D, Percentage of neurons with a significant slope from regression with RT in the given time bin. The same conventions are used as in C.

Figure 8.

Histogram of neurons showing evidence accumulation activity for each task epoch. Dashed line indicates 5%.

Figure 9.

Example neuron with evaluation activity. A–D, Spike-density functions (Gaussian kernel, σ = 20 ms) for correct (A, B) and error (C, D) trials for CONTRA (A, C) and IPSI (B, D) choices. Colors indicate coherence (see inset in A). For each panel, the three plots show activity aligned on onsets of motion stimulus, saccade, and either reward (REW) or error feedback (FDBK), respectively. E, F, Average activity of the example neuron in the 150–400 ms window after reward/error feedback (mean ± SEM) as a function of percentage coherence (E) and reward probability estimated from monkey's behavior (F) in the recording session.

Figure 10.

Population summary of evaluation activity. A, Histogram of neurons with evaluation activity in different task epochs. B, Number of neurons responding significantly differently (solid bars) and not differently (open bars) to outcome-indicating cues (reward onset for correct trials and feedback onset for error trials, Wilcoxon rank-sum test, p < 0.05). CTRL, Neurons without evaluation activity; Stim, Sac, Post, and Rew, neurons with evaluation activity in the corresponding epoch. ALL, Neurons with evaluation activity in any epoch. *The proportion of significant neurons differs from CTRL (χ² test, p < 0.05).

Figure 11.

Bias activity. A, Activity of an example neuron before and during motion viewing on correct trials. For clarity, only 3.2% (blue) and 51.2% (red) coherence are shown. Note that the activity before stimulus onset was different between trials with different final choices (solid vs dashed lines) at 3.2% coherence, but not at 51.2% coherence. B, Predictive index of the example neuron's activity, computed separately for different coherence (colors, as indicated) in sliding 200 ms bins (20 ms steps). Values >0.5 indicate that the neural activity was predictive of the monkey's choice. Activity was aligned to stimulus onset. C, Scatter plot of predictive indices for low- (3.2% and 6.4%) versus high- (25.6% and 51.2%) coherence trials, computed using the average spike rate in the Pre epoch. Red symbols indicate cells with bias activity.

Figure 12.

Caudate activity is task dependent. A, Timeline of the MGS task and epochs used for analysis. B, Comparison of preferred target direction for the dots (left) and MGS (right) tasks. Each row represents the same neuron across panels. Same conventions as Figure 3A. Only neurons that were recorded in both tasks are included. C, Summary of response profiles on the MGS task for neurons with evidence accumulation activity on the dots task. The four panels depict activity of neurons with evidence accumulation activity for four epochs of the dots task, as indicated above. Within each panel, each row represents a single neuron, and each column represents one of the four epochs of the MGS task as in A. White, Activity not different from baseline (Wilcoxon rank-sum test, p ≥ 0.05); black, activity different from baseline, but not direction selective (Wilcoxon rank-sum test, p ≥ 0.05); red, direction selective. D, Summary of response profiles on the MGS task for neurons with evaluation activity on the dots task. The same conventions are used as in C.

Figure 13.

Distribution of neurons along the AP axis. A–D, Histograms of AP locations for neurons relative to the anterior commissure (AC) with decision-related activity (black) and those without decision-related activity (gray) in the Stim (A), Sac (B), Post (C), and Rew (D) epochs. Triangles indicate the median values.