Prior expectation enhances sensorimotor behavior by modulating population tuning and subspace activity in sensory cortex

Abstract

Prior knowledge facilitates our perception and goal-directed behaviors, particularly when sensory input is lacking or noisy. However, the neural mechanisms underlying the improvement in sensorimotor behavior by prior expectations remain unknown. In this study, we examine the neural activity in the middle temporal (MT) area of visual cortex while monkeys perform a smooth pursuit eye movement task with prior expectation of the visual target’s motion direction. Prior expectations discriminately reduce the MT neural responses depending on their preferred directions, when the sensory evidence is weak. This response reduction effectively sharpens neural population direction tuning. Simulations with a realistic MT population demonstrate that sharpening the tuning can explain the biases and variabilities in smooth pursuit, suggesting that neural computations in the sensory area alone can underpin the integration of prior knowledge and sensory evidence. State-space analysis further supports this by revealing neural signals of prior expectations in the MT population activity that correlate with behavioral changes.

Expectation-induced neural modulation in visual area MT accounts for reduced variability in sensorimotor behaviors.

INTRODUCTION

When interacting with our environment, we quickly adapt to its dynamic changes and adjust our actions based on the quality of the sensory information. If the available sensory information is sufficiently precise to guide behaviors, then our actions can depend solely on the sensory input. However, if the sensory information is imprecise, then we combine it with prior knowledge to improve our behavioral responses. In particular, we use the reliability of each piece of information as a weight when combining multiple pieces of information. This precision-weighted information integration explains the fundamental approaches to our interactions with the environment.

The Bayesian inference framework has successfully explained information integration in perceptual decision-making (1–4), multisensory integration (5–9), and sensorimotor behaviors (10–13). The neural implementation of the Bayesian inference has also been demonstrated in studies on sensory and motor functions (5, 14–22). Notably, spontaneous activity in the smooth eye movement region of the frontal eye field (FEF_SEM) represented the prior distribution for the incoming motion speed, whereas the evoked activity reflected a Bayesian estimate of the stimulus speed during smooth pursuit eye movements (14, 23, 24). However, the effect of prior expectations on neural sensory representation has achieved insufficient consensus (25). Human neuroimaging studies have suggested that prior expectations reduce neural activity (26–31) and modulate sensory representation in the early visual cortex (31–33). In contrast, previous single-cell electrophysiological data have shown that prior expectations have no effect on the sensory motion representation in the middle temporal (MT) area of the visual cortex (34). Therefore, although some experimental evidence supports the neural origins of the Bayesian inference, the effects of prior expectations on the cortical sensory neurons are scarce and contradictory.

In this study, we recorded single neuronal activity from area MT of two rhesus monkeys performing a smooth pursuit eye movement task wherein the strength of the sensory motion information and prior knowledge of motion direction were controlled. Consistent with the findings of our previous behavioral study (35), the variation in the pursuit directions across the trials was reduced by prior expectations only when the sensory input was weak and imprecise. The neural recordings indicated that prior expectations systematically reduced the MT neural responses in a manner that sharpened the population direction tuning curve only when the sensory evidence was weak. The simulation of a realistic MT population activity demonstrated that this systematic reduction of the neural responses could account for the reduction in the behavioral variability observed in the initiation of the smooth pursuit eye movement. The integration of prior knowledge with sensory information in area MT and its role in behavioral enhancement was additionally supported by the targeted dimensionality reduction (TDR) analysis of the neural population activity. The neural state of the MT population represented prior information and its temporal evolution accounted for the dynamic modulation of the behavioral variability reduction observed in the smooth pursuit initiation.

RESULTS

In this study, we aimed to identify the neural correlates of prior expectations in area MT using a behavioral paradigm that allowed the direct manipulation of prior information with an identical sensory stimulus (35). We measured the spiking activity of the single neurons in area MT of two rhesus monkeys while performing a smooth pursuit eye movement task (Fig. 1A). Each trial began with a fixation spot appearing at the center of the monitor screen. A random-dot kinematogram appeared after the fixation duration, and all dots of the stimulus moved in a direction within a fixed window for 100 ms (local motion). Following the local motion, the dot patch smoothly moved in the same direction as the local motion, and the monkeys had to maintain their gaze on the visual target during its movement. A liquid droplet was delivered as a reward for the correct behavior at the end of the trial. The expectation of the motion direction was manipulated in two blocks: “wide prior block” and “narrow prior block” (Fig. 1B). Beginning with a randomly selected block, the two prior blocks were alternated. In the wide prior block, the direction of the pursuit target was randomly selected from three widely distributed directions, which were 120° apart. The number of trials in all three directions was the same. In the narrow prior block, the target direction was one of the three narrowly distributed directions, wherein the difference between the central direction and the others was 15°. The central direction was presented twice as often as the other directions. Therefore, the monkeys had prior expectations of the central direction only in the narrow prior block. Notably, the central direction of the narrow prior block, termed the “prior direction,” was also included in the wide prior block to enable a comparison of the eye movements and neural responses between the two blocks under the identical stimulus condition. The prior direction was set by considering the preferred directions of the recorded MT neurons. To change the sensory evidence, we randomly interleaved two different stimulus types, “high contrast” and “low contrast”: These had different luminance contrasts (100% and 12% for monkey A; 100% and 8% for monkey B) and coherence (without and with random-walk noise) of dots in both blocks (Fig. 1C and see Materials and Methods). In addition, in each prior block, the smooth pursuit trials were randomly interrupted using direction tuning trials to measure the direction tuning curves of the neurons being recorded (see Materials and Methods for details). Through direction tuning measurements, neural responses to 12 motion directions (from 0° to 330° in 30° increments) were obtained, and the direction tuning functions of individual neurons were estimated using a Gaussian or circular Gaussian function.

Fig. 1. Experimental design and task performance.

(A) Trial timeline: Each trial begins with monkeys focusing on a central point on the monitor. A random-dot kinematogram then appears, either centrally or off-center, with dots moving in a target direction for 100 ms within an invisible area. The dot patch proceeds to move in the same direction for 500 to 700 ms, while monkeys maintain gaze on the moving patch. The initial fixation duration is randomized at 800, 1300, or 1800 ms. (B) Block design: Each block has different target direction distributions and proportions. Both blocks share a prior direction, with angles between that direction and others being 120° and 15° in the wide and narrow prior blocks, respectively. The number of trials for the prior direction equals that of other directions in the wide prior block but doubles in the narrow prior block. Wide and narrow prior blocks are denoted by red and green fixation points, respectively. (C) Stimulus type: Two random-dot patch types are randomly interleaved. High-contrast stimuli have 100% luminance contrast and dot coherence, while low-contrast stimuli have 12% (monkey A) or 8% (monkey B) luminance contrast and random-walk noise in dot motion. (D and E) Mean SD of pursuit directions for prior direction targets: Black and red bars represent mean SD for high-contrast targets in wide and narrow prior blocks, while gray and yellow bars indicate mean SD for low-contrast targets in the respective blocks. Error bars denote standard error of the mean (SEM). ***P < 0.0005.

Prior expectation reduces the variability of pursuit eye movements

Our previous study (35) demonstrated that prior expectation of the motion direction reduces the variation in the pursuit directions for moving stimuli. We reconfirmed the effect of the direction prior on pursuit eye movements. To compare the variability in the pursuit eye movements in the same direction between the wide and narrow prior blocks, we calculated the SD of the pursuit directions for the prior direction in each block during the open-loop period [100 ms from pursuit onset (36)] to exclude any impact of feedback signals (see Materials and Methods). Figure 1 (D and E) shows the mean SD of the pursuit directions from the two monkeys (59 days’ data for monkey A and 67 days’ data for monkey B). In monkey A, the mean SD of the narrow prior block was smaller than that of the wide prior block under both high- and low-contrast conditions [mean SD in the wide and narrow prior blocks was 10.28° versus 9.72° for high-contrast stimuli, paired t test, t(58) = 3.72, P = 4.46 × 10⁻⁴; and 16.20° versus 13.64° for low-contrast stimuli, paired t test, t(58) = 12.70, P = 2.18 × 10⁻¹⁸]. In monkey B, the difference in the SD between the two blocks was significant only for low-contrast stimuli [mean SD in the wide and narrow prior blocks: 6.94° versus 7.00° for the high-contrast stimuli, paired t test, t(66) = −0.61, P = 0.54; and 11.66° versus 9.83° for low-contrast stimuli, paired t test, t(66) = 6.94, P = 2.33 × 10⁻⁹]. A two-way analysis of variance (ANOVA) revealed that the interaction between the effects of the stimulus contrast and prior block on the pursuit SD as well as each effect was significant [main effect in monkey A, P = 1.18 × 10⁻³⁸ (stimulus contrast) and P = 1.02 × 10⁻⁶ (prior block); interaction effect in monkey A, F(1,232) = 10.41 and P = 0.0014; main effect in monkey B, P = 6.90 × 10⁻¹⁷ (stimulus contrast) and P = 0.031 (prior block); interaction effect in monkey B, F(1, 264) = 4.98 and P = 0.026]. These results demonstrate that prior expectation of the motion direction decreases the variation in the pursuit directions, particularly when the sensory motion is weak and imprecise, which is consistent with the prediction of the Bayesian inference (35). We also fitted the smooth pursuit eye movements to the Bayesian observer model, confirming that the model accounted for the direction expectation induced pursuit biases and the reduction in SD. Overall, monkey A’s sensory representation was less precise than that of monkey B, as evidenced by a likelihood function with a larger SD (fig. S1, K and L). This result explains why only monkey A exhibited a prior-induced reduction in behavioral variability under high-contrast conditions. The weaker motion representation of monkey A made the effect of the prior more noticeable under high-contrast conditions (see fig. S1, M and N, for schematics illustrating the changes and interactions among prior probability, posterior probability, and likelihood functions).

Other basic properties of the smooth pursuit eye movements, when comparing the two prior conditions (including example eye velocity traces), are summarized in fig. S1. The effect of prior expectation on pursuit latency and speed varied between the monkeys. Monkey A exhibited quicker (low latency) and faster pursuit (high speed), particularly under low-contrast conditions, while monkey B displayed more delayed (high latency) and slower pursuit (low speed) under both contrast conditions (fig. S1, C to F). Acceleration revealed more frequent and higher-amplitude catch-up saccades under wide prior conditions than under narrow ones (fig. S1, I and J), which is expected considering that catch-up saccades would depend on the size of the sensory error (the discrepancy between sensory motion and eye velocity) at the end of the open-loop period. The presence of higher and more frequent catch-up saccades under a wide prior condition, regardless of stimulus contrast, also demonstrated the consistent influence of prior expectation in both monkeys.

Prior expectation systematically reduces MT neural responses only when sensory input is weak and imprecise

Recent human functional magnetic resonance imaging studies have reported that the prior expectation of a stimulus feature reduces neural responses in the primary visual cortex (29, 31) and enhances neural representations in early visual areas (31, 32). A more recent magnetoencephalography study supported the role of expectation in the perceptual processes by demonstrating that the expectations modulate the neural representation in early sensory processing (33). In this study, we tested whether the responses of the MT neurons were modulated by prior expectations. To determine the neuronal correlations underlying the effect of prior expectation on the pursuit, we recorded 257 well-isolated single neuronal activities (137 for monkey A and 120 for monkey B) from area MT while the monkeys performed the task. The peristimulus time histograms (PSTHs) in Fig. 2 (A and B) show the firing rate of the MT neurons in the prior direction as a function of time. When the stimulus contrast was low, the firing rate in the narrow prior block was lower than that in the wide prior block in the initial neural responses of both monkeys [cluster-based permutation test (37), 10,000 permutations, α = 0.05; monkey A, 61 to 129 ms; monkey B, 101 to 199 ms from the local motion onset]. However, no common decrease in the firing rate was observed between the two monkeys when the stimulus contrast was high. We primarily focused on early neural responses because it is difficult to determine whether the neural response reduction occurring after the open-loop period (up to 100 ms after pursuit onset or 200 ms after motion onset) is induced by prior expectation or by changes in retinal input that occur once the eyes begin to move. To further quantify the reduction in the initial neural responses, we calculated the firing rates during the initial 100 ms from the spike latency of each neuron in response to the local motion. The insets in Fig. 2 (A and B) depict the mean firing rates of the MT neurons in the time window under the two prior conditions. The early MT neuronal responses after the spike latency were smaller in the narrow prior block than in the wide prior block only when the stimulus contrast was low [mean firing rate of the wide and narrow prior blocks in monkey A was 43.37 versus 42.64 for the high-contrast case, paired t test, t(135) = 1.82, P = 0.07; and 25.21 versus 23.17 for the low-contrast case, paired t test, t(110) = 4.53, P = 1.51 × 10⁻⁵; that in monkey B: 49.86 versus 49.38 for the high-contrast case, paired t test, t(116) = 1.15, P = 0.25; and 25.45 versus 23.98 for the low-contrast case, paired t test, t(109) = 3.35, P = 0.001]. These results indicate that prior expectation of the motion direction reduces the responses of the MT neurons according to the strength of the motion. When the sensory motion is weak and imprecise, the effect of prior expectation appears to be more pronounced.

Fig. 2. Effect of prior expectation on the firing rate of MT neurons.

(A to C) Each colored PSTH shows the firing rate of the MT neurons for the prior direction as a function of time relative to the local motion onset in each condition (black: wide prior, high contrast; red: narrow prior, high contrast; gray: wide prior, low contrast; yellow: narrow prior, low contrast). The bar graphs show the mean firing rates during the initial 100-ms time window from the spike latency. *P < 0.05. (D to F) Correlation between the difference in the prior and preferred directions (Δθ) and the firing rate ratios (FR ratios) in the narrow prior block to that in the wide prior block. The lines show the correlation coefficient over time relative to the local motion onset with a ±30-ms window and 5-ms timestep; the red, blue, and black lines are from monkey A, monkey B, and the combined dataset of monkeys A and B, respectively. (G to I) Correlation between Δθ and FR ratio in the time window of 55 to 115 ms from the local motion onset in the combined dataset. Each point in the plot represents one MT neuron. The black solid lines show the linear regression between Δθ and the log FR ratio. (C, F, and I) The MT neuronal responses for low-contrast stimuli on the days where the SD in the pursuit directions in the narrow prior block are significantly smaller than those in wide prior block and the responses on the other days with no significant difference in the SD between the two prior blocks.

The small and uniform decrease in neural responses may not account for the decrease in behavioral variability. However, if the response reduction depends on the difference between the preferred directions of the neurons and prior direction, then it can enhance the neural representation of the motion direction by modulating the population direction tuning function. Prior expectation–induced reduction in the MT neuronal responses indeed varied according to the relationship between the prior direction and the preferred direction of each neuron. We calculated the ratio of the firing rate in the narrow prior block to that in the wide prior block. We then estimated the correlation coefficient (Spearman’s rho) between the firing rate ratio (FR ratio) and the difference in the prior and preferred directions (Δθ). The correlation from each monkey and that from the combined data are plotted as a function of time relative to the local motion onset (Fig. 2, D and E). The correlation did not differ from zero when the stimulus contrast was high (Fig. 2D). However, when the stimulus contrast was low, the correlation decreased after the stimulus appeared, and it was significantly lower than zero between 80 and 100 ms from the local motion onset in the combined data [Fig. 2E, time windows for calculation: ±30 ms at each time point of 0 to 180 ms in 5-ms intervals, Spearman’s correlation, false discovery rate (FDR)–corrected (38), α = 0.05]. We used Spearman’s rho to demonstrate the relationship between the FR ratio and Δθ, as nonparametric measures were more reliable and not affected by outliers. Figure 2 (G and H) shows the relationship between the FR ratio and Δθ at 85 ± 30 ms (Spearman’s correlation, ρ = −0.01, P = 0.87 for high contrast; ρ = −0.284, P = 0.00059 for low contrast). This means that as the preferred direction of an MT neuron deviated further from the prior direction, the neuron’s response to the prior direction decreased more when the sensory input was weak. This relationship remained significant for neurons whose preferred direction was less than 90° apart from the prior direction, although the overall correlation was weaker (ρ = −0.18, P = 0.0477 for low contrast). When we divided the neurons into three groups based on the size of Δθ (30% of total neurons with small Δθ, another 30% with medium Δθ, and the last 30% with large Δθ, using a 100-ms firing rate window from spike latency), a similar result to the correlation analysis was observed. The firing rate differences between the prior conditions were the largest when Δθ was substantial (fig. S2).

The animals’ performance on the pursuit task varied from day to day, with monkeys using prior knowledge of the motion direction more effectively in some sessions than in others. To examine whether neural modulation depended on the monkeys’ usage of prior knowledge, we divided the experimental sessions into two groups based on whether prior expectations significantly reduced the variability of the pursuit direction. We compared the systematic reduction of neural activity between the two groups: sessions with statistically significant behavioral effects due to prior expectation versus sessions without significant behavioral effects (number of neurons for significant versus nonsignificant pursuit direction SD reduction under the low-contrast conditions: n = 62 for 29 days versus n = 75 for 30 days in monkey A; n = 83 for 35 days versus n = 37 for 32 days in monkey B; two-sample F test, α = 0.05).

When the pursuit direction variance was significantly reduced by prior expectation, the responses of the MT neurons also decreased, and the response reduction depended on Δθ, as observed when using the entire dataset. However, when the reduction in pursuit direction variance was not significant, the response reduction still occurred in both cases, but the correlation between the reduction and Δθ was much weaker than before [Fig. 2C: mean firing rate in wide versus narrow prior block of monkey A: 30.13 versus 27.93 on the days with significant behavioral effect, paired t test, t(61) = 2.73, P = 0.009; and 21.32 versus 19.41 on the days with nonsignificant behavioral effect, paired t test, t(74) = 3.83, P = 3.04 × 10⁻⁴; that of monkey B: 20.52 versus 18.88 on the days with significant behavioral effect, paired t test, t(82) = 2.99, P = 0.004; and 35.56 versus 34.48 on the days with nonsignificant behavioral effect, paired t test, t(36) = 1.51, P = 0.14; Fig. 2, F and I: Spearman’s correlation between FR ratio and Δθ in 85 ± 30 ms from the local motion onset: ρ = −0.39, P = 0.001 on the days with significant behavioral effect and ρ = −0.18, P = 0.11 on the days with nonsignificant behavioral effect].

This relationship suggests that the systematic reduction of neural responses was controlled by the strength of the prior expectation used by the monkeys. It also indicates that the decrease in pursuit direction variability due to prior expectation may be driven by the systematic response reduction of cortical sensory neurons.

Decoded direction information from the simulated MT responses can account for the behavioral modulation by prior expectation

The systematic reduction in neuronal responses, which depends on the difference between the prior direction and the preferred directions of MT neurons, can change the shape of the population direction tuning curve in area MT. If the prior expectation is congruent with the stimulus motion, then it can sharpen the population direction tuning curve and enhance the neural representation of the motion direction (Fig. 3A). To test whether the observed response reduction has a noticeable impact on the quality of the sensory neural representation and the resulting behavioral performance, we simulated neuronal activity using realistic properties obtained from our MT recordings. Using the estimated parameters of the circular Gaussian functions (fig. S3), which were fitted to recorded MT neural responses for obtaining direction tuning functions, we simulated 3600 neuron responses whose preferred directions were randomly selected from a uniform distribution bounded by −180° and 180° (see Materials and Methods for details). We obtained measurements of the tuning shape, average firing rates, Fano factors, and interneuronal correlations from our neuronal recordings. These elements were incorporated into the simulation to ensure that the simulated neurons exhibited the realistic neural response properties we observed in the recordings (see Materials and Methods for further details). Subsequently, the effect of prior expectations on the neural activity was modeled by applying slopes of a linear regression between Δθ (preferred direction–prior direction) and the log FR ratio (narrow/wide prior) that was estimated from the recordings (regression coefficient at 85 ± 30 ms from the local motion onset: −1.06 × 10⁻⁴ for the high-contrast target, P = 0.60; −0.0025 for the low-contrast target, P = 1.01 × 10⁻⁶). The simulated population responses for the high-contrast stimuli did not differ considerably between the two prior conditions because the slope of regression was close to zero (but still negative). However, the simulated population tuning curves for the low-contrast stimuli became narrower under the narrow prior condition because of the negative relationship between Δθ and the FR ratio. In addition to the Spearman’s correlation, we also obtained the coefficients of the linear regression with confidence intervals across time (fig. S4). The temporal pattern of the regression coefficient changes was similar to that of the correlation coefficient. The estimated regression coefficients were significantly different from zero around the time window of interest (85 ± 30 ms from the local motion onset). Figure 3 (B and C) shows examples of the simulated responses for high- and low-contrast conditions, respectively, when the target and prior directions are 0°.

Fig. 3. In silico simulations to decode population direction information.

(A) Schematic of population tuning curve change in area MT by prior expectation: Black and red curves display the population direction tuning of model MT neurons in wide and narrow prior blocks when the target and prior directions are both 0°. When sensory evidence is weak, prior expectations proportionally reduce MT neuronal responses according to the difference between the prior direction and each neuron’s preferred direction, sharpening the population tuning curve. (B and C) Examples of simulated MT population responses: Thin lines represent the firing rate of each model neuron in a single trial under high/low-contrast and wide/narrow prior conditions. Thick curves depict mean population responses across trials. (D) Support vector machine (SVM) performance: SVM’s ability to discriminate between two directions as a function of direction difference (Δθ) is shown by colored circles fitted to a cumulative Gaussian function (dot-dashed line). (E and F) Estimation of direction SD ratio: The ratio between narrow and wide prior blocks in experimental and simulated data is represented by gray circles for individual sessions’ smooth pursuit directions in the two monkeys, and black circles for the mean SD ratios. Black lines show the SD of black circles. Light red crosses indicate the SD ratio estimated from population responses in each simulation using a population vector decoder (PVD) (F, left) or maximum likelihood estimation method (F, right), with thick red crosses representing the mean SD ratio.

Using these simulated MT neural responses, we tested whether the neural direction discrimination could be improved by the prior factor. Neural population responses to two different motion directions were simulated and a binary linear support vector machine (SVM) discriminated the two directions, in which one was 0° and the other was a direction from 0° to 10° with step sizes of 0.5° (see Materials and Methods for details). In this simulation, the target and prior directions were identical, and the prior factor was applied to each direction separately such that the population tuning curve became centered on each stimulus direction (Fig. 3, B and C). Therefore, the prior factor only influenced the level of noise in the neural direction representation because it only sharpened the population direction tuning response to a given motion direction without shifting the peak. When the difference between the two directions was small, the SVM classifier could not discriminate between the directions (Fig. 3D, direction difference < 2°). However, its performance rapidly increased as the direction difference increased. The performance of the SVM classifier was better under the high-contrast condition than under the low-contrast condition [mean of the cumulative Gaussian function under high- and low-contrast conditions of wide prior block: 2.97 versus 4.45, paired t test, t(99) = −42.26, P = 3.77 × 10⁻⁶⁵; that of narrow prior block: 2.92 versus 3.57, paired t test, t(99) = −19.79, P = 3.52 × 10⁻³⁶]. However, the effect of the prior factor under the high-contrast condition was negligible [mean of the cumulative Gaussian function for high-contrast stimuli: 2.97 (wide) versus 2.92 (narrow), paired t test, t(99) = 1.66, P = 0.1]. In contrast, the effect of the prior factor was significant under the low-contrast condition. When the classifier was trained and tested on the simulated neural responses under the low-contrast, narrow prior condition, the neural direction discrimination was significantly better than that when the classifier was trained under the wide prior condition [mean of the cumulative Gaussian function for low-contrast stimuli: 4.45 (wide) versus 3.57 (narrow), paired t test, t(99) = 26.04, P = 4.28 × 10⁻⁴⁶]. These findings suggest that the systematic response reduction observed in the MT neuronal activity to low-contrast stimuli might have an effect on enhancing the neural motion direction representation, probably by reducing the population-level neural noise.

The SVM demonstrated that prior expectations improve the signal-to-noise ratio in the population-level neural representation of the motion direction. To determine how much of the behavioral enhancement by prior expectation can be accounted for by the modulation of the population direction tuning in area MT, we compared the population-level neural variability (i.e., variance in neural direction representation) with the behavioral variability (i.e., variance in pursuit direction). For this comparison, a population vector decoder (PVD) (39) and maximum likelihood estimate (MLE) (40–42) were used to directly measure the trial-by-trial variability of the neural direction estimates (see Materials and Methods for details). The direction in each trial was decoded from the simulated population response in a 0° motion direction, and the SD of the direction estimates was calculated across trials. To estimate the extent to which the direction variability was reduced by prior expectation, we computed the ratio of the SDs in the narrow and wide prior blocks (SD_{narrow prior}/SD_{wide prior}). The mean SD ratio of the direction estimates from the simulations was smaller than 1, which suggests that the simulated SD in the narrow prior block was smaller than that in the wide prior block and the effect was stronger when the stimulus contrast was low [Fig. 3, E and F; mean SD for high-contrast stimuli: 0.994 and 0.993, one-sample t test, t(99) = 2.74 × 10⁴ and 384.16, P = 0 and 6.39 × 10⁻¹⁵⁹; mean SD for low-contrast stimuli: 0.87 and 0.80, one-sample t test, t(99) = 1.37 × 10³ and 157.58, P = 1.70 × 10⁻²¹³ and 1.12 × 10⁻¹²⁰ from PVD and MLE, respectively]. This was consistent with the findings from the behavioral data [Fig. 3, E and F; mean SD for high-contrast stimuli: 0.99, one-sample t test, t(123) = 89.92, P = 1.83 × 10⁻¹¹⁵; mean SD for low-contrast stimuli: 0.87, one-sample t test, t(123) = 75.47, P = 4.00 × 10⁻¹⁰⁶]. Notably, the SD ratio of the directions estimated by the PVD was almost the same as the experimental SD ratio (experimental versus simulated SD ratio in the high-contrast case: 0.9863 versus 0.9936, pooled t test, t = −0.60, P = 0.55; those in low-contrast case: 0.8677 versus 0.8721, pooled t test, t = −0.34, P = 0.73). This close correspondence between the experimental and simulated results suggests that the modulation of the shape of the population direction tuning curve can fully account for the reduction in the behavioral variation in smooth pursuit initiation.

As demonstrated before (35), prior expectations introduced an increase in the behavioral bias and a decrease in the behavioral variability during the pursuit task, especially when the sensory input was weak and imprecise. The directions of the smooth pursuit eye movements to the outer target directions were skewed toward the prior direction: The ratio between the angular difference of the two outer target directions and that of the corresponding pursuit directions was significantly smaller than 1 under the low-contrast conditions [fig. S5A; 0.997 and 0.801 for high- and low-contrast stimuli, paired t test, t(42) = −0.25 and −9.66, P = 0.81 and 1.82 × 10⁻¹⁵]. To test whether the bias can also be explained by the expectation-induced modulations in the population tuning, we simulated population responses in the ±15° motion direction when the prior direction was 0° and decoded the average directions using the PVD and MLE, respectively (see Materials and Methods for details). Consistent with the experimental results, the simulation showed that the estimated directions were biased toward the prior direction, and this effect was significantly stronger when the stimulus contrast was low (fig. S5B; the ratio between the two differences for high-contrast stimuli: 0.994 and 0.987, one-sample t test, t(99) = −244.59 and −57.47, P = 1.57 × 10⁻¹³⁹ and 7.54 × 10⁻⁷⁸; that for low-contrast stimuli: 0.84 and 0.75, one-sample t test, t(99) = −261.12 and −1.12 × 10³, P = 2.45 × 10⁻¹⁴² and 8.29 × 10⁻²⁰⁵ from PVD and MLE, respectively). These results demonstrate that the systematic change in the population response of the MT neurons by prior expectations can explain the bias increase and variability reduction in the pursuit directions. In addition, this further suggests that sharpening the population direction tuning in area MT might be sufficient to explain the Bayesian inference observed in smooth pursuit eye movements.

Systematic reduction in neuronal responses occurs only under an active behaving condition

As mentioned earlier, prior expectations enhance pursuit behavior by sharpening the shape of population neural direction tuning. However, the systematic response reduction could arise from passive response modulation due to repetitive presentation of similar stimuli. That is, these observations might stem from neural adaptation (43) rather than prior expectation, although the two are closely correlated. To address this possibility, we measured the direction tuning responses of the MT neurons in the middle of each prior block. Even if the same motion stimuli were presented, the monkeys did not have to use prior knowledge because the animals passively fixed their eye gazes on the central fixation point. In contrast, changes in the passive properties of individual neurons that were induced by the different statistical distributions of the motion stimuli across the blocks could be revealed. To measure the direction tuning, we used identical visual stimuli as the pursuit targets and placed them in the center of the receptive fields of the neurons. The directions were randomly selected from 12 directions (0°, 30°, …, 300°, 330°) in each presentation. To obtain the mean direction tuning curve across all neurons from the two monkeys, we pooled the direction tuning responses by aligning them relative to the preferred direction of each neuron. Subsequently, we fit the realigned mean neural responses to the Gaussian function and compared the averaged direction tuning function between the wide and narrow prior blocks. We could not observe a significant difference in the mean direction tunings between the two blocks (Fig. 4, 95% bootstrap confidence intervals of the fitted Gaussian function parameters a_i, b_i, ${\mathrm{\theta }}_i^{\mathrm{pref}}$, and σ_i of wide and narrow prior blocks in the combined data, monkey A, and monkey B were summarized in table S1). Moreover, the responses were not significantly different across the block conditions [Fig. 4A, FDR-corrected P values (q values) > 0.4 under high-contrast condition and q > 0.6 under low-contrast condition for the combined dataset; Fig. 4B, q > 0.3 under high-contrast condition and q > 0.3 under low-contrast condition for monkey A; Fig. 4C, q > 0.7 under high-contrast condition and q > 0.5 under low-contrast condition for monkey B; Fig. 4D, q > 0.9 under high-contrast condition and q > 0.5 under low-contrast condition for the large Δθ neurons; and Fig. 4E, q > 0.3 under high-contrast condition and q > 0.9 under low-contrast condition for the small Δθ neurons; see table S2 for individual values]. In addition, the tuning function of the individual neurons was estimated by fitting the Gaussian or circular Gaussian functions (selecting a better function between the two for each neuron). The amplitudes and half-widths of the tuning curves were not different across the blocks under both high- and low-contrast conditions [paired t test, mean tuning amplitude in wide and narrow prior blocks was 55.48 versus 55.05 for high contrast, t(164) = 0.86, P = 0.39, and 32.03 versus 31.88 for low contrast, t(164) = 0.32, P = 0.75; mean tuning half-width in wide and narrow priors was 85.11 versus 86.71 for high contrast, t(164) = −0.82, P = 0.41, and 87.88 versus 87.08 for low contrast, t(164) = 0.18, P = 0.86].

Fig. 4. Direction tuning responses of MT neurons in passive fixation task.

(A) Mean direction tuning across all MT neurons in the combined data from the two monkeys. The direction tuning responses of individual neurons are realigned relative to each neuron’s preferred directions (filled circles) and fitted to a Gaussian function (dash-dotted lines). (B and C) Mean direction tuning of MT neurons in each monkey. (D and E) Mean direction tuning of top and bottom 30% neurons based on the size of the difference between each neuron’s preferred direction and prior direction in the combined data.

To further test whether the change in the direction tuning function of the MT neurons occurred only when the preferred direction of each neuron was far from the prior direction, we divided the neurons into two groups: one with larger direction discrepancies between each neuron’s preferred direction and the prior direction and the other with smaller discrepancies (i.e., top and bottom 30% of neurons based on the direction discrepancies). We then averaged the direction tunings of the neurons in each group. As shown in the results from the entire population, the direction tunings of the neurons in both groups were not different between the wide and narrow prior blocks [Fig. 4, D and E; top 30% neurons with larger Δθ: mean tuning amplitude in wide and narrow priors was 39.37 versus 38.82 for high contrast, t(45) = 0.98, P = 0.33, and 20.41 versus 21.43 for low contrast, t(45) = −1.68, P = 0.1; mean tuning half-width in wide and narrow priors was 85.42 versus 89.30 for high contrast, t(45) = −1.11, P = 0.27, and 96.47 versus 89.97 for low contrast, t(45) = 0.45, P = 0.65; bottom 30% neurons with smaller Δθ: mean tuning amplitude in wide and narrow priors was 63.60 versus 62.27 for high contrast, t(45) = 1.16, P = 0.25, and 33.71 versus 32.97 for low contrast, t(45) = 1.04, P = 0.31; mean tuning half-width in wide and narrow priors was 84.96 versus 82.16 for high contrast, t(45) = 1.53, P = 0.13, and 87.21 versus 86.44 for low contrast, t(45) = 0.43, P = 0.67]. This suggests that the systematic response reduction did not result from the passive neural response changes induced by the repetitive presentation of the motion stimulus. These results exclude the adaptation accounts and demonstrate that the systematic reduction in the neuronal response only appears when prior knowledge of the motion direction is used. We further evaluated the effect of passive adaptation by observing the initial trials after the block change. When we only considered the first 20 trials after the block changes, we still observed a significant reduction of both the behavioral SD and neural response (fig. S6). This suggests that the effect of prior expectations occurs fairly fast; therefore, the behavioral and neural modulations are not likely to originate from the passive adaptation.

Prior expectation is represented in the neural subspaces

The monkeys alternated their use of prior expectations between trial blocks, enabling the expectation-related signal to influence the activity of MT neurons even before the appearance of the pursuit target in the narrow prior block. However, when this possibility was examined in the average PSTHs during the pursuit and fixation tasks, prior expectations did not seem to strongly affect spontaneous activities [Figs. 2 (A and B) for both monkeys and 5 (A and D) for monkey A and fig. S7 (A and D) for monkey B]. The mean spontaneous activity was marginally increased by prior expectations during the pursuit task only for monkey B (mean firing rate from −400 to 0 ms in wide and narrow prior blocks: 9.23 versus 9.43, P = 0.26 in monkey A; 8.86 versus 9.20, P = 0.036 in monkey B). The mean firing rate of the MT neurons showed significant differences only after the local motion onset in both monkeys (Fig. 2, A and B). Although a consistent effect of prior expectations on the average spontaneous neural activity was not observed across the two monkeys, it could potentially exist within the activity patterns of the neural population. Recent studies have demonstrated that context, sensory inputs, and perceptual choices can be represented in task-relevant dimensions of population responses in the prefrontal cortex (44, 45).

Fig. 5. Latent dynamics of MT population responses during pursuit and fixation task.

(A and D) PSTHs relative to the local motion onset and mean spontaneous activities of the recorded MT neurons (inset), in each prior block during the pursuit task (A) and fixation task (D). (B and E) Autocorrelation matrix of regression coefficients for prior blocks during the pursuit task (B) and fixation task (E). (C and F) Temporal trajectories of average population responses in the task-related subspace consisting of prior and contrast axes during the pursuit task (C) and fixation task (F). Trajectories of sponteneous activity (−400 to 0 ms in the pursuit task and −100 to 0 ms in the fixation task) are shown in the inset plots, and dash-dotted lines highlighted these duration. (G) SDs of pursuit directions with high-contrast (left) and low-contrast (right) stimuli as a function of time relative to the local motion onset. The red and yellow thick lines indicate that the SD differences between the two blocks are significant [cluster-based permutation test (37), 10,000 permutations, P < 0.05]. (H) Differences in the population responses projected on the prior axis of the subspace between the wide and narrow prior blocks as a function of time (solid lines), and the differences of the pursuit direction SDs between the two priors as a function of time (dashed lines). The blue and magenta lines indicate the high- and low-contrast conditions, respectively. Temporal coherences between the subspace differences and SD differences with different latencies of the SD differences from the subspace differences are shown in the inset plots. The thick top line in the inset indicates that the FDR-corrected P-values are less than 0.05. (I) Differences in the prior axis–projected population responses between the two prior blocks during the fixation task (averaged across 12 motion direction conditions).

To determine whether the MT population encodes task variables, including prior expectations and sensory information, we used the TDR method (44). Since only the pursuit targets of the prior directions were presented in both the wide and narrow prior blocks, we limited our subsequent analyses to the neural responses to the prior directions (refer to fig. S8 for the analysis that included responses to all stimulus directions). For the population analysis, we constructed population responses by pooling all the units recorded primarily in separate sessions. We estimated the linear regression coefficients for the prior expectation and stimulus contrast by explaining each neuron’s individual responses across trials as a linear combination of the task variables (prior expectation and stimulus contrast) and then estimating the coefficients associated with the task variables (see Materials and Methods for details). To denoise the estimated regression coefficients, we projected them onto the subspace spanned by the dominant 12 principal components (PCs) of the population activity and then reprojected the denoised coefficients onto the original basis. The regression coefficients for the prior, which reveal the dependencies of MT neural responses on prior block changes, were temporally correlated during the spontaneous activity period (Fig. 5B for monkey A and figs. S7B for monkey B and S8, A and B, for regression coefficients for the contrast). This temporal correlation indicates that the MT population activity retains prior information even before the appearance of the pursuit targets.

Next, we defined the task-related axes by time-averaging the regression coefficients for prior and contrast and then orthogonalizing them. We projected the averaged population responses across trials onto these axes, generating neural trajectories in the task-related subspace (see Materials and Methods for details). The neural trajectories illustrate the temporal evolution of the average population response during the pursuit task trial. The projected population response moved in the direction of high contrast (upward) and wide prior (rightward) after the local motion onset, driven by sensory information from the stimulus (Fig. 5, C and F, and fig. S7, C and F). The general direction of these trajectory changes results from the sensory inputs: Sensory stimuli contain luminance contrast information, which ranges from 0 to 100% (upward changes; no visual stimulus, 8 or 12% contrast, and 100% contrast), and the appearance of the sensory stimulus resolves the uncertainty of the upcoming motion direction, thereby reducing the need for expectations of the motion direction (rightward changes). Notably, the population responses in the wide and narrow prior blocks were distinguishable along the prior axis (Fig. 5C and fig. S7C for monkeys A and B, respectively), with clear separation already observed during spontaneous activity (insets in Fig. 5C and fig. S7C, dotted black circles). We quantified this separation by calculating the difference in the prior-axis projections of the neural trajectories; prior-axis projection of wide prior population response − prior-axis projection of narrow prior population response. This difference in prior-axis projection was significantly maintained throughout the trial, including the fixation period, only in the pursuit task but not in the fixation task (Fig. 5, H and I, for monkey A and fig. S7, H and I, for monkey B). The mean difference in prior axis–projected population responses between the wide and narrow prior blocks during the pursuit task was 1.04 and 1.18 for high- and low-contrast cases in monkey A and 0.60 and 0.72 in monkey B, and during the spontaneous activity period, it was 0.42 and 0.60 for high- and low-contrast cases in monkey A and 0.48 and 0.35 in monkey B (two-sided permutation test, P < 0.05; see Materials and Methods for details).

The latent dynamics along the prior axis were functionally relevant to the pursuit behaviors, consistent with the simulation results. We calculated the SDs of the pursuit directions over time to obtain the dynamics of behavioral variability during the pursuit trial and averaged them across the behavioral sessions (Fig. 5G for monkey A and fig. S7G for monkey B). The SDs rapidly decreased during the open-loop period (approximately 100 to 200 ms from the local motion onset). The reduction in the SD in the narrow prior block was stronger than that in the wide prior block, particularly when the contrast of the sensory stimulus was low.

To determine whether the differences in the prior-axis projections of the subspace trajectories were related to behavioral variance reduction, we compared the temporal changes of the prior-axis projections of the neural subspace differences with those of the pursuit SD differences. Both the differences in subspace and in SD were at their highest shortly after the onset of local motion (Fig. 5H for monkey A and fig. S7H for monkey B). Only in the low-contrast condition did the subspace differences exhibit a high temporal correlation with the SD differences, and the subspace differences consistently appeared ahead of the SD differences. A significant correlation of 0.5 between the two indicated that the SD differences appeared 60 to 120 ms later than the subspace differences in monkey A and 40 to 80 ms later in monkey B (insets in Fig. 5H and fig. S7H, FDR-corrected, α = 0.05). The temporal evolution of the behavioral variability difference was similar to that of the prior-axis projection difference across prior blocks, indicating that the temporal changes of the prior input were similarly expressed in both behavior and population neural responses. This result additionally suggests that the neural modulation of the MT population responses by prior expectations may contribute to a reduction in the variability in oculomotor behavior.

Neural subspace modulation was specific to the pursuit task. When the same population analysis was applied to the neural population data from the fixation task, the modulation of the population neural spontaneous activity along the prior axis of the subspace disappeared. Regression coefficients for the prior were not correlated across time for almost the entire duration (Fig. 5E for monkey A and fig. S7E for monkey B). Furthermore, the trajectories within the neural subspace were not separable between the wide and narrow prior blocks before the visual stimulus appeared, thus demonstrating the absence of prior expectation–related input to area MT (inset in Fig. 5F for monkey A and inset in fig. S7F for monkey B). There was no difference in the neural trajectories projected to the prior axis between the wide and narrow prior blocks during the fixation task (two-sided permutation test, P > 0.1; the test was done for the prior direction only). These results suggest that the expectation signal was present in the neural subspace of the MT activity only when the monkeys were engaged in using prior knowledge in the task.

Expectation-induced neural subspace modulation in the spontaneous activity suggested a possibility of differential activity modulation across the neurons. Similar to the evoked responses, the modulation of the spontaneous activity by the prior expectation could depend on the Δθ. To test this possibility, we measured Spearman’s correlation between Δθ values and log FR ratios, as in the evoked responses. Even if it was fairly weak, we found a significant negative correlation in the spontaneous activity during the pursuit task (ρ = −0.15, P = 0.04). We could not find a significant correlation in the spontaneous activity during the fixation task (ρ = −0.08, P = 0.3): Therefore, this negative correlation was task specific (fig. S9). We also divided the neurons by the difference between each neuron’s preferred direction and the prior direction and observed the difference in the spontaneous activity between the prior conditions. Consistent with the correlation analysis, the spontaneous activity under the narrow prior condition was significantly higher than that under the wide prior condition when the neurons’ preferred directions were close to the prior condition (Δθ < 45°; fig. S9, A to D). These results suggested that the differential modulation of the spontaneous activity that depends on Δθ might underlie the neural subspace modulation by the prior expectation.

Notably, the modulation of the spontaneous neural activity was not induced by anticipatory eye movements during fixation. Moreover, the modulation of the evoked neural responses by the prior expectation was not the result of the eye movements preceding the local motion onset. We measured the eye speeds during fixation and compared them across the prior conditions. The fixational eye speeds were not different across conditions in monkey A, but they were higher under the narrow prior condition in monkey B (fig. S9, G and J). To test the possibility of these fixational eye movements (including anticipatory saccades and pursuits) affecting the upcoming sensory motion information (anticipatory eye movements would affect the visual sensory motion information), we calculated the projection of the spontaneous eye velocity to the direction of the prior expectation on a given session (fig. S9, H and K). The spontaneous eye velocities toward the prior direction were not different across the prior conditions in monkey A; however, these were significantly higher under the narrow prior condition in monkey B. This result suggested that monkey B has anticipatory eye movements toward the prior direction under the narrow prior condition. However, this anticipatory eye movement would effectively decrease the sensory motion toward the prior direction when a pursuit target appeared. Subsequently, the neural response of the neurons whose preferred direction matched the prior direction would decrease, and the responses of the neurons whose preferred directions were far from the prior direction would increase, which would result in the exact opposite effect from what we found. In addition, the neural effect in the current study was consistent across the two monkeys, even if the fixational eye movement pattern was different between the two. This strongly suggested that the neural modulation that we observed was not due to anticipatory fixational eye movements.

DISCUSSION

In this study, prior expectations reduced the variation in the pursuit directions and discriminately reduced the responses of the MT neurons, especially when the sensory evidence was weak, depending on their preferred directions. Using data-based simulations and multiple decoders (SVM, PVD, and MLE), we demonstrated that the systematic response reduction improves the neural representation of the motion direction by sharpening the population direction tuning in area MT, which accounts for behavioral improvement. In addition, the TDR analysis of the population neural responses demonstrated the existence of the prior expectation signal in the neural subspace of spontaneous activity and the functional relevance of the subspace dynamics to the behavioral variability reduction observed in pursuit initiation.

Neural origins of prior expectation

The Bayesian observer model has successfully revealed the integration of prior expectation and sensory information in smooth pursuit eye movements (35): however, the corresponding neural areas and mechanisms for this inference are largely unknown. Notably, whether the sensory area only represents the likelihood function of the visual input or also represents the prior information and the posterior probability distribution is unknown. The hierarchical Bayesian inference framework for visual processing provides a theoretical ground for understanding the neural mechanisms underlying the interplay between the sensory and top-down signals (46–49). Within this framework, higher-order brain regions may suppress lower-order neural responses, particularly when they are “incongruent” with prior expectations (48). Thus, bottom-up sensory signals are reduced in such a manner as to increase the signal-to-noise ratio. Recent human neuroimaging studies have provided empirical support for this hypothesis by showing the expectation-driven modulation of neural activity and representation in the visual cortex (29, 31, 32). In addition, a recent electroencephalogram study showed that motion expectation signals reaching the sensory regions are relevant to the expectation-induced behavioral bias in smooth pursuit eye movements (50). However, these neuroimaging data cannot explain how the visual system improves the sensory representation by silencing the incongruent responses: In addition, the neural responses that are incongruent with the expectations are ambiguous. Here, we propose a neural mechanism in the visual cortex that elucidates how prior expectation operates, involving a systematic reduction in the neural responses leading to sharpened population direction tuning. The simulation revealed that prior expectations can change the pursuit direction bias and variability via the systematic reduction in the MT neural responses. These results underpin the Bayesian brain hypothesis by showing that the sensory area may represent not only the likelihood function but also the posterior distribution to which prior information is applied.

A possible alternative framework for sensory processing is the traditional feedforward model. Under this framework, the integration of sensory evidence and prior information occur in the parietal and frontal areas but not in the sensory areas (51, 52). A neurophysiological study by Rao et al. (34) supported this perspective: The responses of neurons in the macaque lateral intraparietal area were affected by prior expectations of the stimulus motion, whereas almost no change was observed in the MT neuronal responses during a direction discrimination task using saccadic eye movements. However, our results of the reduced MT responses by prior expectations contrasted these findings. These different results may be attributed to at least two possible factors. First, the previous research and this study used different eye movement systems (saccades and pursuits) for behavioral tasks. Although saccadic and pursuit eye movements share some common parts of the oculomotor network, they have largely disparate phylogenies and functions (53). Therefore, although area MT is involved in both saccades and pursuits, the influence of cognitive modulation on the MT neural activity may differ depending on the eye movement system. Second, the two studies used different behavioral paradigms to control the monkeys’ prior expectations of the motion direction of the stimuli. Rao et al. (34) used an arrow cue for “transient” top-down modulation, whereas we used “cumulative” learning based on experience-dependent prior knowledge from previous trials. These different strategies for developing an expectation of the motion direction may have resulted in different neural mechanisms.

In our results, the prior expectation signal was present in the single neuronal responses in area MT after the pursuit target appeared and was also represented as the pattern of the population responses even before the local motion onset. The TDR analysis (44, 54) clearly showed the effect of prior expectations on the spontaneous population activity (Fig. 5, B and C, for monkey A and fig. S7, B and C, for monkey B), and this effect was task specific (absence of the effect in the fixation task; Fig. 5, E and F, for monkey A and fig. S7, E and F, for monkey B). These results strongly suggest the existence of the “prior” inputs in the activity of the MT neurons. Notably, the prior expectation signal that appeared in the spontaneous activity was maintained during smooth pursuit initiation, regardless of the strength of the sensory inputs (neural subspace differences between the wide and narrow prior blocks: Fig. 5, C and H, for monkey A and fig. S7, C and H, for monkey B). Neurons in area MT may receive top-down expectation-related inputs from other brain regions, and these signals may be dynamically integrated with feedforward sensory inputs evoked by the pursuit target. Therefore, the neural activity in area MT may represent the posterior probability distribution by integrating the prior and sensory inputs during the presentation of the pursuit target. Both univariate (unit) and multivariate (population) analyses suggest the reliability-weighted integration of these two inputs by showing that the expectation-evoked neural changes accounted for the behavioral modulation in pursuit initiation when the sensory evidence is weak (Figs. 3, D to F, and 5H). However, because most neurons were not simultaneously recorded, our analyses could not directly show the modulation of both the population tuning and population response pattern by prior expectations. Future work should aim to identify the contribution of the population-level representations of the sensory and cognitive information to behavior through the direct analysis of the trial-by-trial relationship between the population activity and behavior.

Which brain regions provide this task-specific expectation signal to area MT? A recent study (14) demonstrated that FEF_SEM represents the prior expectation of the speed through spontaneous activity; a prior expectation of faster speed results in higher spontaneous activity during fixation before the visual motion onset. If the spontaneous activity in the FEF_SEM also represents the prior expectation of the motion direction, then the source of prior information could be FEF_SEM, although the manner of representing the directional expectation could be different from the simple gain modulation. Moreover, the precision-weighted integration between the prior and likelihood functions of the motion speed was represented in the evoked responses of the FEF_SEM. Because both area MT and FEF_SEM represented prior-likelihood integration, the two sites appear to have an essential role in the effect of prior expectations on smooth pursuit eye movements while interacting with each other. A previous study reported that prior expectations of the direction of a visual stimulus increase the functional connectivity between area MT and prefrontal cortex (55). Investigating the role of each region and the interaction between the two in implementing the Bayesian inference during perceptual decision-making and sensorimotor behavior is an important direction for future research.

Expectation and attention

Expectation and attention have a similar facilitatory effect on visual perception; expected and attended stimuli are more easily detected and recognized than unexpected or unattended stimuli. Therefore, the effects of the expectation on perceptual decisions and sensorimotor behavior have often coalesced with those of attention in the empirical literature. In numerous cases, changes in perception and behavior due to expectations can also be explained by attention, and the distinction between these effects is unclear. For example, the behavioral and neural effects of the expected stimuli could be caused by decreased attention and vice versa. Although the mechanisms overlap and interact, neither expectation nor attention is solely responsible for the influences of the two on visual perception. To fully understand expectation and attention, they must be conceptually dissociated (56–59). Although expectation and attention were not strictly separated in our task design, the behavioral results were not due to spatial attention as the pursuit targets were always presented at the same position (fovea) between the two prior blocks. Unlike spatial attention, feature-based attention (i.e., attending to the direction of motion) affecting the perceptual behavior during our task could be concerning. Determining whether the results were from prior expectation or feature-based attention from behavioral data alone is challenging; however, the neural results may provide insight into this.

In terms of the neural modulation, attention increases the responsivity of the sensory neurons (60), whereas expectations decrease the neural activity in the sensory areas (29, 31, 61). In area MT, spatial (62, 63) and feature-based attention (64, 65) increases the gain of the direction tuning curves of the MT neurons without changing the tuning bandwidths. However, the effects of expectations on these neurons remain largely unknown. In this study, prior expectations reduced the visual responses of the MT neurons, which is consistent with the results of previous neuroimaging studies, and the reduction in the neural activity could contribute to improving the behavioral reliability. This supports the observation that the reduced responses were not due to lower attention but a higher expectation because less attention to stimuli cannot enhance behavioral performance.

Furthermore, previous attention studies reported that behavioral performance could be improved by reducing the trial-by-trial variability of the single neuronal responses in the visual cortex (66–68). Thus, if prior expectation and attention operate on the same neural mechanism, then prior expectations may reduce the variation in the eye movements by modulating the spike count variability of the MT neuronal responses. We measured the mean-normalized variance (i.e., Fano factor) of the spike counts and compared it across different conditions. A transient decrease in the Fano factor occurred after the local motion onset in all conditions (wide/narrow prior × high/low contrast), and the decrease was larger in the high contrast than in the low-contrast conditions (fig. S10, A and B), which was consistent with previous results (69). However, the Fano factor did not differ between prior conditions. In monkey A, when the stimulus contrast was high, there was a period in which the Fano factor in the narrow prior block was lower than that in the wide prior block [cluster-based permutation test (37), 10000 permutations, P < 0.05]. However, this effect was observed after the eye began to move; therefore, it could be caused by the difference in eye movements. In addition, this effect was negligible in monkey B. These results suggest that the prior expectation of the motion direction does not significantly change the variability of the single-neuronal spiking.

In addition to single-neuron variability, the modulation of trial-by-trial spike count correlations between neurons can influence perceptual performance. Previous studies have shown that the attention-induced reduction in interneuronal correlation is closely related to improved behavioral performance (66, 68, 70, 71). In a similar vein, it is possible that prior expectations might reduce interneuronal correlations, thereby contributing to decreased noise at the neural population level. To examine this possibility, we measured and compared trial-by-trial correlations between pairs of neurons across the wide and narrow prior blocks. We found that the pairwise spike count correlations between neurons did not vary between these prior conditions (see fig. S10, C to F). These results indicate that neither the trial-by-trial spike count variability of individual neurons nor the trial-by-trial correlation between pairs of neurons can account for the reduction in behavioral variability due to prior expectations. Thus, the neural mechanisms of expectation and attention during visual perception and oculomotor responses may differ, at least superficially. However, it is important to acknowledge that we cannot completely rule out the possibility that prior expectations could modulate interneuronal correlations, as we did not have enough neuronal pairs to reach a definitive conclusion (refer to Materials and Methods for the number of pairs).

Despite finding no indications of feature attention’s involvement in the neural modulation we observed, we must not discount this possibility completely. The differences could be quantitative in nature; some latent, underlying neural mechanisms might be shared between attention and expectations. However, they could also operate under entirely distinct neural mechanisms. A well-designed experimental paradigm and computational models would be necessary to properly dissociate expectation effects from attention effects and systematically explore the similarities and differences between the two mechanisms.

Expectation and neural adaptation

Previous studies have reported the adaptation of MT neurons to the repeated motion direction of visual stimuli (43, 72–74); neural adaptation reduced the amplitude and width of the direction tuning curves of the MT neurons, and the effects were greatest when the preferred direction of the MT neurons and the direction of the adapting stimulus matched. In this study, the pursuit targets for the prior direction were presented twice as often as those of the other directions in the narrow prior block. However, the neural response reduction to the prior direction differed from the passive response modulation of the neural adaptation for the following reasons. First, the direction tuning curves of the MT neurons in the narrow prior block were similar to those in the wide prior block when the animals passively fixed their eye gazes at the central fixation point (Fig. 4). Second, contrary to what was expected from neural adaptation, the effect was stronger when the prior direction differed from the preferred direction. Third, the amount of response reduction by the prior expectation did not depend on the extent to which the receptive field and pursuit target overlapped (fig. S11). Last, the effect appeared relatively fast. The behavioral variability and neural response reduction happened in 20 trials after the prior condition changed from wide to narrow (fig. S6). These differences imply that the reduced responses of the MT neurons may not have been due to prolonged exposure to repeated visual motion but rather because of cognitive feedback signaling, which is different from the simple mechanistic reaction of the neurons to adjust constant sensory input.

Effect of prior expectation under high-contrast condition

We did not observe a strong effect of the prior block when the sensory stimulus was strong and reliable (under the high-contrast condition). This result is consistent with the prediction of the Bayesian inference model, as it reflects reliability-weighted information integration. However, we still observed an effect of the prior even under high-contrast conditions, albeit weak. For instance, we found a significant reduction in behavioral variability due to the prior expectation in monkey A when the pursuit target had 100% contrast (Fig. 1D). This result could be attributed to the less reliable motion representation of monkey A, as inferred from the results of fitting the behavioral data to the Bayesian observer model (fig. S1, K and L); the SDs of monkey A’s likelihood function were larger than those of monkey B. In our neural recordings, we did not find evidence that reflects these individual differences under high-contrast conditions. This could be due to the difficulty in detecting subtle differences using population direction tuning and decoding techniques; the negative relationship between the change in the preferred direction (Δθ) and FR ratio may not have achieved the desired sensitivity to reveal the effect under the high-contrast condition and individual monkey differences. However, using a more sensitive multivariate analysis of the population neural activity pattern (TDR analysis), we demonstrated the presence of the prior signal in the activity of MT neurons, regardless of the pursuit target’s contrast status, whether it was low or high (even in spontaneous activity). Future studies using advanced recording techniques, experimental designs, computational methods, or a combination of these approaches may provide a clearer answer to the unanswered questions.

MATERIALS AND METHODS

Two 10-year-old adult male rhesus monkeys (Macaca mulatta) weighing 9 to 11 kg were used for the neurophysiological experiments. All research protocols were approved by the Sungkyunkwan University Institutional Animal Care and Use Committee (SKKUIACUC2017-06-10-1). Before the experiments, we performed two separate surgeries. A head holder was implanted on the skull for the head restraint, after which a cylindrical chamber fabricated from PolyEther Ether Ketone (PEEK) was implanted on the skull close to the lunate sulcus for an angled approach to area MT. During the surgeries, each monkey was under isoflurane anesthesia, while antibiotics and analgesics were administered postoperatively to minimize infection and pain.

Visual stimuli and behavioral paradigm

Visual stimuli were presented on a gamma-corrected 24″ Cathode-Ray Tube (CRT) monitor (HP1230, 1600 × 1200 pixels, 85-Hz vertical refresh rate). The monitor was placed 570 mm from the animal, and it covered 38.67° by 29.49° of the horizontal and vertical visual field, respectively. The background on which the visual stimuli were presented was gray with a luminance level of 36.8 cd/m² (luminance range was 0 to 73.68 cd/m²). The presentation of the visual stimuli and recording of the eye movement data were controlled using a real-time data acquisition system (Maestro version 3.3.11). A custom-built photodiode system was used to ensure the accurate timing of the visual stimuli.

The two monkeys were trained in a smooth pursuit eye movement task (Fig. 1A). The pursuit target was a random dot kinematogram, and its size in each day’s experiment was determined as one of 4° × 4°, 8° × 8°, or 12° × 12° depending on the location and size of the receptive fields of the recorded MT neurons. Each trial began when the animals fixated their eyes on a small dot (fixation spot) at the center of the monitor screen. After a randomized fixation duration (800, 1300, or 1800 ms), the dot patch appeared at the center of the screen or 1 to 2° displaced from the center to the opposite direction of the target direction. Subsequently, a local motion, whereby all the dots inside the invisible circular window moved in the target direction at a given speed but the invisible window did not move, was implemented. Following the local motion, both the dots and windows moved together at the same speed and in the same direction as the local motion for 500 to 700 ms. If the animals maintained their eyes at the center of the pursuit target within a 4° window during the target movements, then they were rewarded with a drop of water or juice.

To control the animals’ prior expectation of the motion direction of the pursuit target, we used two types of blocks: wide and narrow prior blocks (Fig. 1B). In the wide prior block, the fixation spot was red, and the pursuit target direction was randomly and evenly selected from three directions, which were 120° apart. Therefore, the animals’ expectations for the incoming motion direction would be widely distributed across the directions. Conversely, in the narrow prior block, the fixation spot was green, and the direction of the pursuit target was one of the three narrowly distributed directions (a central direction and ±15° of the direction). The central direction was presented twice as often as the other two directions. Therefore, the animals tended to develop a prior expectation for this central motion direction or the expectation would be narrowly distributed around the central direction. Therefore, we termed this direction the prior direction. The prior direction was set by considering the preferred directions of the recorded neurons. Notably, identical prior directions were included in both the wide and narrow prior blocks to compare the effect of the prior expectations on the behavioral and neural responses to identical sensory stimuli. The narrow and wide prior blocks consisted of 252 and 378 trials, respectively. The first block in each day’s experiment was randomly selected and the two prior blocks were alternated. We typically collected four blocks for each category, resulting in 2520 trials in total.

Further, the precision of the sensory evidence (sensory stimulus) was controlled by randomly selecting a pursuit target stimulus from the two stimulus types in each trial: One was a random-dot kinematogram with 100% luminance contrast (high contrast), and the other was a random-dot kinematogram with 12 or 8% (for monkeys A and B, respectively) luminance contrast (low contrast). To further reduce the precision of the motion direction in the low-contrast case, a random-walk direction noise (75) was included in the stimulus, such that the directions of all dots were randomly selected from uniform distribution bounded by ±60° of the target direction (Fig. 1C).

Data collection and analysis

The horizontal and vertical eye positions were separately recorded at a sampling rate of 1 kHz using an infrared video tracking system (EyeLink 1000 Plus, SR Research Ltd.). A low-pass Butterworth filter with an order of two and a cutoff frequency of 20 Hz was applied on the horizontal and vertical components of the eye position. Subsequently, the filtered eye positions were differentiated to obtain the horizontal and vertical eye velocities. To focus on the initiation of the smooth pursuit eye movements and remove the influence of the saccadic eye movements (saccades) on the behavioral responses, the trials with saccades occuring within a time window between −100 and 250 ms from the onset of the pursuit target were removed (we did not apply this exclusion criterion and included all trials in the neural data analysis). Next, the eye velocity traces were rotated such that their average direction was 45°. The rotated eye velocity during the open-loop period [100 ms from pursuit onset (36)] was decomposed by horizontal and vertical templates (76, 77), which were the average of each rotated eye velocity component between −20 and 100 ms from the average pursuit latency (mean pursuit onset for high- and low-contrast stimuli: 76 and 123 ms in monkey A and 57 and 105 ms in monkey B). The optimal pursuit latency and scaling factors were estimated through the sliding and scaling of the two templates using the least-squares method and the nonlinear optimization with mesh adaptive direct search (NOMAD) algorithm (78).

A single-electrode or an eight-channel laminar probe (Multitrode Type Ι, Thomas RECORDING) with impedances from 0.5 to 1 megohm (at 1 kHz) was used to record the spikes and local field potentials (LFPs) in the MT using the Motorized Electrode Manipulator System (Thomas RECORDING). Extracellular electrical signals from the electrode were high pass–filtered at a cutoff frequency of 150 Hz, digitized at a sampling rate of 40 kHz for action potentials, low pass–filtered at a cutoff frequency of 170 Hz, and digitized at a sampling rate of 1 kHz for LFPs (OmniPlex). The approximate receptive field location and size of a well-isolated MT neuron were measured using a hand-controlled visual stimulus while the monkeys fixated their eye gaze on a stationary spot at the center of the monitor. The MT neurons whose receptive fields were near the fovea were analyzed as the pursuit targets were presented at or near the fovea.

To isolate the single-unit responses, we sorted the spikes offline (Offline Sorter, Plexon) using principal components analysis and the waveforms of the recorded neural activities. For further analysis, we included only the neurons whose spikes formed distinct clusters in the principal component space. In addition, we checked and removed the sorting errors by inspecting the time stamps of the sorted spikes with a time resolution of 1 ms. Thus, we obtained 137 neurons from 59 days of recordings of monkey A and 120 neurons from 67 days of recordings of monkey B. For behavioral analysis, only days with more than 50 trials for each condition (four conditions: wide/narrow prior block and high/low-contrast stimuli) were used (59 days for monkey A and 65 days for monkey B).

Spike count analysis

To investigate the effect of prior expectations on neural responses, we compared the responses of MT neurons to the prior direction stimulus in wide and narrow prior blocks. We constructed a PSTH for each prior condition and smoothed it using a 10-ms SD Gaussian filter (Fig. 2, A and B). In addition, we counted the spikes in the 100-ms duration time window from the spike latency of each neuron and compared them between the prior blocks (insets in Fig. 2, A and B). To estimate each neuron’s spike latency, we averaged the spiking responses across all repetitions of the given stimulus contrast (high or low) over time, and the latency was determined through visual inspection of the mean PSTH. If the neuronal latency could not be visually determined because of a lack of distinct stimulus-related responses, the neuron was excluded from the analysis. Therefore, we used data from 136 and 111 neurons for the high- and low-contrast conditions, respectively, in monkey A and from 117 and 110 neurons for each contrast condition in monkey B. To calculate the mean-normalized variance (variance/mean, Fano factor) of the spike counts over time, we used a 100-ms (±50 ms) sliding time window with 20-ms steps (fig. S10, A and B). For the in silico simulations, we imposed the Fano factor estimated from the time period between 0 and 200 ms after the local motion onset (Fig. 3).

Direction tuning measurement and estimation

Before the smooth pursuit eye movement task, the direction and speed tuning of the isolated MT neurons were estimated in a separate direction tuning task. To assess the direction tuning of the neurons, the stimulus motion was presented in 12 directions (0 to 330° in 30° increments) at the preferred speed of the MT neurons while the animals fixated their eye gaze on a central stationary spot. A circular patch of random dots with 100% coherence served as the visual stimulus, and the location and size of the visual stimulus were set according to the cells’ receptive field properties. Each trial began with a fixation duration between 600 and 1200 ms, and then six different motion stimuli were presented for 256 ms with 300 ms of interposed stationary epochs. For the direction tuning curve, the spike counts for 256 ms, after 50 ms from the stimulus onsets, were fitted to a Gaussian or circular Gaussian function

$${\mathrm{MT}}_i(\mathrm{\theta })=a_i{e}^{[-\frac{{(\mathrm{\theta }-{\mathrm{\theta }}_i^{\mathrm{pref}})}^2}{2{({\mathrm{\sigma }}_i)}^2}-1]}+b_i(\mathrm{Gaussian})$$ (1)

$${\mathrm{MT}}_i(\mathrm{\theta })=c_i{e}^{\left\{d_i[\cos (\mathrm{\theta }-{\mathrm{\theta }}_i^{\mathrm{pref}})-1]\right\}}(\text{circular Gaussian})$$ (2)

where θ is the 12 motion directions of the visual stimulus; a_i, b_i, ${\mathrm{\theta }}_i^{\mathrm{pref}}$, and σ_i are the parameters of the Gaussian function of ith MT neuron, where ${\mathrm{\theta }}_i^{\mathrm{pref}}$and σ_i denote the mean and SD of the Gaussian function, respectively; c_i and d_i are the parameters of the circular Gaussian function of ith MT neuron; and ${\mathrm{\theta }}_i^{\mathrm{pref}}$indicates the preferred direction of the MT neuron. The direction tuning responses for the different luminance contrasts were obtained by randomly interleaving the high-contrast (100% luminance contrast for both monkeys) and low-contrast (luminance contrast of 12% for monkey A and 8% for monkey B) patches as the visual stimuli. For speed tuning of the neurons, the motions at nine speeds (0, 1, 2, 4, 8, 16, 32, 64, and 128°/s) with the preferred direction of the neurons were presented. The speed tuning responses were fitted to a log Gaussian function (79). If the direction tuning exhibited a bimodal shape or the estimated preferred speed was extremely high or low, then the orientation tuning was also quantified to identify whether the neuron had an orientation selectivity rather than direction selectivity. Six orientations (0°, 30°, 60°, 90°, 120°, and 150°) were used for the orientation tuning task and excluded the neurons when their orientation tuning responses were more robust than the direction tuning responses.

In addition, the direction tuning responses of the MT neurons in the middle of the smooth pursuit task were intermittently measured by randomly inserting trials of the direction tuning task. The visual stimuli for the direction tuning measurement were identical to the pursuit targets, but they were placed and remained in the center of the receptive fields to evoke strong MT neuronal responses. To obtain the average tuning responses of the neurons, we aligned the direction tuning responses of individual neurons to each neuron’s preferred direction such that all preferred directions were at 0° (Fig. 4). We excluded neurons from this analysis if their preferred direction was not well defined or if tuning measurements in either prior block were missing. We used data from 107 and 106 neurons for the high- and low-contrast conditions from monkey A and from 87 and 86 neurons for high- and low-contrast conditions from monkey B. We estimated the bootstrap confidence interval of the Gaussian function parameters for the average direction tuning curve using subsets of trials for each neuron (100 resamplings). In addition, to characterize each neuron’s tuning properties, we fitted the directional responses of a neuron measured during the pursuit task to a Gaussian or circular Gaussian function. We excluded neurons from all analyses using these fitting functions if the fitted tuning curve explained less than 50% of the response variance. The preferred direction of each neuron was determined on the basis of the tuning function for the high-contrast stimulus in both prior blocks, except when one of the prior block data was missing, in which case only one prior block was used. We used data from 97 neurons from monkey A and 90 neurons from monkey B. The estimated preferred directions obtained from this process were used in the analyses aimed at observing the relationship between the prior-induced neural response reduction and the difference in the preferred direction of each neuron and the prior direction, denoted by Δθ.

Interneuronal spike count correlation

Trial-by-trial spike count correlations between pairs of MT neurons were calculated using Spearman’s correlation coefficient (figs. S10, C to F). We used 100-ms (±50 ms) sliding time windows from −200 to 400 ms after the local motion onset with a step size of 20 ms. Pairs of neurons were included in the sample only if the number of trials for the prior direction was more than 50 in both the wide and narrow prior blocks. This criterion was applied separately for each stimulus contrast condition. This resulted in the analysis of 165 (high contrast) and 99 (low contrast) pairs from monkey A and 123 (high contrast) and 117 (low contrast) pairs from monkey B.

Relationship between the difference of preferred and prior directions and the FR ratio

To investigate the relationship between Δθ and the reduction of MT neural responses due to prior expectations, we used the Spearman’s correlation and linear regression. We computed Spearman’s correlation and linear regression coefficients between Δθ and the log FR ratio of the two prior blocks for each 60-ms (±30 ms) sliding time window from the local motion onset to 180 ms with 5-ms time intervals (Fig. 2, D to I). In this analysis, neurons with spike latencies shorter than 100 ms were included to minimize the impact of oculomotor behaviors on sensory responses. With this additional criterion, 181 (n = 95 for monkey A and 86 for monkey B) and 144 (n = 75 for monkey A and 69 for monkey B) neurons were retained for the high- and low-contrast cases, respectively.

In addition, to examine whether the relationship between Δθ and the FR ratio is associated with the reduction in pursuit direction SD, we conducted the same analyses (Spearman’s correlation and linear regression) separately for the sessions with statistically significant and nonsignificant behavioral effects (two-sample F test, α = 0.05). For the sessions with a significant behavioral effect, 140 neurons (69 from monkey A and 71 from monkey B) were used under the high-contrast condition, and 64 neurons (35 from monkey A and 29 from monkey B) were used under the low-contrast condition. For the sessions without a significant behavioral effect, 41 neurons (26 from monkey A and 15 from monkey B) were used under the high-contrast condition, and 80 neurons (40 from monkey A and 40 from monkey B) were used under the low-contrast condition.

Simulation and decoding of MT population responses

For the in silico simulations, the firing rates of the MT neurons in response to the stimulus motion were modeled as a circular Gaussian function (Eq. 2) based on our experimental dataset (72). The tuning parameters were estimated from the combined dataset between the wide and narrow prior blocks when data from both prior conditions were available, or from one prior block data, for each contrast stimulus (80 neurons from monkey A and 87 neurons from monkey B were used for both high- and low-contrast conditions). Each circular Gaussian function parameter for the responses of the ith model MT neuron ($c_i^{\mathrm{\prime }}$ and $d_i^{\mathrm{\prime }}$) was randomly selected from a gamma distribution fitted to the estimated parameters of the direction tuning functions of the recorded MT neurons using the gamfit function in MATLAB

$$c_i^{\mathrm{\prime }}\sim \text{Γ}(k_c,{\mathrm{\theta }}_c),d_i^{\mathrm{\prime }}\sim \text{Γ}(k_d,{\mathrm{\theta }}_d)$$ (3)

where k_c is the shape parameter, θ_c is the scale parameter of the gamma distribution of c, and k_d and θ_d are the shape and scale parameters of the gamma distribution of d, respectively (fig. S3). The MT neuronal responses to the high- and low-contrast stimuli were simulated (k_c= 1.74 and 1.28; θ_c = 36.83 and 31.16; k_d = 1.30 and 1.18; θ_d = 1.19 and 0.73 for the high- and low-contrast stimuli, respectively). The preferred directions (${\mathrm{\theta }}_i^{\mathrm{pref}}$) ranged from −179° to 180° in 1° increment, and each preferred direction was assigned to 10 simulated neurons. Thus, the model totally included 3600 (360 × 10) neurons for every simulation. Trial-by-trial variation in the individual neuronal responses and trial-by-trial correlation between the neurons were simulated using the mean Fano factor and mean interneuronal correlation over 200 ms from the local motion onset estimated from the MT recordings. Because we could not find any evidence of significant changes in the Fano factors and interneuronal correlations across prior conditions, we averaged the Fano factors and interneuronal correlation coefficients from the wide and narrow prior blocks in each contrast case (Fano factor for high contrast: 1.11, low contrast: 1.37; interneuronal correlation for high contrast: 0.02, low contrast: 0.06). The correlation between the neuronal responses was simulated using the Cholesky decomposition method with a constant interneuronal correlation coefficient for each contrast condition (80). To model the reduction effect of prior expectations on the neuronal responses, we multiplied the simulated MT responses by the exponential of the regression coefficient between Δθ (|preferred direction – prior direction|) and the log FR ratio (narrow/wide prior) computed from the experimental data (regression slope at 85 ± 30 ms from the local motion onset for high contrast, −0.0003, and for low contrast, −0.0025; see Fig. 2, G and H).

$${\mathrm{MT}}_i^{\mathrm{high},\mathrm{narrow}}(\mathrm{\theta })=e^{-0.0003\times \mid {\mathrm{\theta }}_i^{\mathrm{pref}}-{\mathrm{\theta }}^{\mathrm{prior}}\mid }\times {\mathrm{MT}}_i^{\mathrm{high},\mathrm{wide}}(\mathrm{\theta })$$ (4)

$${\mathrm{MT}}_i^{\mathrm{low},\mathrm{narrow}}(\mathrm{\theta })=e^{-0.0025\times \mid {\mathrm{\theta }}_i^{\mathrm{pref}}-{\mathrm{\theta }}^{\mathrm{prior}}\mid }\times {\mathrm{MT}}_i^{\mathrm{low},\mathrm{wide}}(\mathrm{\theta })$$ (5)

where θ is the direction of the target motion, ${\mathrm{\theta }}_i^{\mathrm{pref}}$ is the preferred direction of the ith neuron, θ^prior is the expected direction (the prior direction), and ${\mathrm{MT}}_i^{\mathrm{contrast},\mathrm{prior}}$ is the average firing rate of the ith neuron to the target direction in each contrast and prior condition, from Eq. 1.

Support vector machine

The SVM classifiers were trained with 80% of the model MT population to discriminate two motion directions using the fitclinear function in MATLAB. With the remaining 20% of the population neural activity, the trained classifier predicted one of two directions, which were 0° and a direction between 0° and 10° (step sizes of 0.5°). We simulated 200 trials of the neuronal population responses to each motion direction and repeated each pair of simulations 100 times to validate the predictive performance of the SVM classifier.

Population vector decoder

Population vector averaging was used to estimate the direction information represented in the neuronal population activity. Using the simulated MT neuronal responses, the PVD predicted the direction of the stimulus motion as follows

$${\mathrm{\theta }}^{\mathrm{\prime }}=\frac{{\displaystyle \sum _{i=1}^n}{\mathrm{\theta }}_i^{\mathrm{pref}}{\mathrm{MT}}_i(\mathrm{\theta })}{{\displaystyle \sum _{i=1}^n}{\mathrm{MT}}_i(\mathrm{\theta })}$$ (6)

where θ is the direction of motion of the visual stimulus, ${\mathrm{\theta }}_i^{\mathrm{pref}}$ is the preferred direction of the ith model MT neuron, MT_i(θ) is the response of the ith neuron obtained from Eq. 1. The target direction θ was set to 0° for the prior direction (Fig. 3E) or ±15° for the outer directions (fig. S5B). To compensate for the estimation error resulting from asymmetries in contributing neural population in PVD, the preferred directions of the model MT neurons ranged from −194° to 165° or from −164° to 195° (instead of ranging from −179° to 180°) when the target direction was −15° or 15°, respectively. We simulated 200 trials and computed the SD of the predicted directions (θ^′) in each simulation. This procedure was repeated 100 times.

Maximum likelihood estimation

The simulated MT responses were also decoded using maximum likelihood estimation (40–42). The log likelihood of the motion direction θ_x can be expressed as

$$L_{\mathrm{\theta }}=k{\displaystyle \sum _{i=1}^n}{n}_i\cos (\mathrm{\theta }-{\mathrm{\theta }}_i^{\mathrm{pref}})$$ (7)

where k is the concentration parameter of the circular Gaussian function, which is the tuning function of the neurons, and n_i is the spike counts of the ith MT neuron in response to the direction θ. The log likelihood of any direction of motion can be computed from any other two nondegenerate log likelihoods, L_θ1 and L_θ2

$$\left[\begin{array}{c}{L}_{{\mathrm{\theta }}_1}\\ L_{{\mathrm{\theta }}_2}\end{array}\right]=\left[\begin{array}{cc}\cos ({\mathrm{\theta }}_1)& \sin ({\mathrm{\theta }}_1)\\ \cos ({\mathrm{\theta }}_2)& \sin ({\mathrm{\theta }}_2)\end{array}\right]\left[\begin{array}{c}{\displaystyle \sum _{i=1}^N}{n}_i\cos ({\mathrm{\theta }}_i^{\mathrm{pref}})\\ {\displaystyle \sum _{i=1}^N}{n}_i\sin ({\mathrm{\theta }}_i^{\mathrm{pref}})\end{array}\right]$$ (8)

From this equation, the log likelihood of any other direction θ₃ can be computed as follows

$$L_{{\mathrm{\theta }}_3}=[\cos ({\mathrm{\theta }}_3)\sin ({\mathrm{\theta }}_3)]{\left[\begin{array}{cc}\cos ({\mathrm{\theta }}_1)& \sin ({\mathrm{\theta }}_1)\\ \cos ({\mathrm{\theta }}_2)& \sin ({\mathrm{\theta }}_2)\end{array}\right]}^{-1}\left[\begin{array}{c}{L}_{{\mathrm{\theta }}_1}\\ L_{{\mathrm{\theta }}_2}\end{array}\right]$$ (9)

In each trial, we calculated the log likelihoods for 0° and 45°, L_θ0° and L_θ45°, as L_θ1 and L_θ2 and then computed the log likelihoods of all other directions between −179° and 180° using L_θ0° and L_θ45°. The direction with the largest log likelihood was used as an estimate of the stimulus motion, indicating the direction representation of the population response.

Targeted dimensionality reduction

To understand the effects of prior expectations on the MT neural activity at the population level, we used the TDR method. A detailed explanation of TDR can be found in Mante et al. (44). We first constructed the population activity by pooling the responses of all the MT neurons across conditions (stimulus contrast, prior block, and target direction) in each monkey (pursuit task: 137 neurons from monkey A and 120 neurons from monkey B; fixation task: 82/86 neurons in the high/low-contrast cases from monkey A and 83/80 neurons in the high/low-contrast cases from monkey B). In the pursuit data, although the prior and outer directions (±15° and ±120° from the prior direction) varied in individual recording sessions, the relative relationships between the prior and outer directions were the same. Therefore, a pseudo population was composed by aligning each neuron’s responses relative to the prior direction, where each neuron’s preferred direction was redefined depending on the difference between the preferred and prior directions.

During the pursuit task, targets in the prior directions were presented in both the wide and narrow prior blocks, but other targets in the outer directions were presented either in the wide or narrow prior blocks. Therefore, we used linear regression with the neural responses to only the prior direction to describe the population responses as a linear combination of task variables (stimulus contrast and prior)

$$r_{i,t}(j)={\mathrm{\beta }}_{i,t}(1)\mathrm{contrast}(j)+{\mathrm{\beta }}_{i,t}(2)\mathrm{prior}(j)+{\mathrm{\beta }}_{i,t}(3)$$ (10)

where r_i,t(j) is the z-scored response of neuron i at time t on trial j (the mean and SD are computed from the neuron’s responses across all trials and times); contrast(j) is the stimulus type on trial j (1, high contrast; 0, low contrast); prior(j) is the block type on trial j (1, wide prior block; 0, narrow prior block); β_i,t(1) and β_i,t(2) are the regression coefficients that reflect the extent to which the trial-by-trial firing rate of neuron i at time t depends on the corresponding task variable, contrast and prior, respectively; β_i,t(3) is the regression coefficient responsible for the trial-by-trial variation in the firing rates due to time rather than the task variables.

During the fixation task, all stimulus motions in 12 directions (0°, 30°, …, 300°, 330°) were presented in both prior blocks across all sessions. Therefore, the MT neural responses to all motion directions were used to compute the regression coefficients for stimulus contrast and prior.

Subsequently, to denoise the regression coefficients, we initially conducted a principal components analysis on the neural population responses. The neural data matrix consisted of cells, time, and task variable conditions. Since we used two task variables (contrast and prior) and each variable had two possible cases (high versus low contrasts and wide versus narrow priors), there were four task variable conditions. We projected them into the PC space composed of the first 12 PCs:

$$\begin{array}{c}\text{Denoised regression coefficients in the subspace}(12\mathrm{PCs},\mathrm{time},\text{task variable})=\\ \text{Projection matrix}(12\mathrm{PCs},\mathrm{cell})\times \text{Regression coefficient matrix}(\mathrm{cell},\mathrm{time},\text{task variable})\end{array}$$

Afterward, to represent the regression coefficients in the original basis, we projected the subspace coefficients back into the original space:

$$\begin{array}{c}\text{Denoised regression coefficients in the original space}(\mathrm{cell},\mathrm{time},\text{task variable})=\\ \text{Transposed projection matrix}(\mathrm{cell},12\mathrm{PCs})\times \text{Denoised regression coefficients in the subspace}(12\mathrm{PCs},\mathrm{time},\text{task variable})\end{array}$$

Next, we defined the time-invariant regression vector for each task variable by taking the regression coefficients at the time interval where the norm of the regression vector was the maximum among all other time intervals. The time intervals with the maximum norm in the pursuit task were 81 to 100 ms from local motion onset for the prior and 61 to 80 ms for the contrast in monkey A, and 61 to 80 ms for both prior and contrast in monkey B; those in fixation task were 101 to 120 ms for the prior and 61 to 80 ms for the contrast in monkey A, and 61 to 80 ms for the prior and contrast in monkey B. The resultant time-invariant regression vector for each task variable is now one-dimensional vector, composed of the number of cells. Last, the time-invariant regression vectors for all task variables were orthogonalized with QR decomposition to be used as task-related subspace axes that independently accounted for the trial-by-trial variance of the population response due to task variables. The average population responses were projected onto these axes to investigate the temporal dynamics of the population responses in the task-related neural subspace.

In addition, we performed the TDR analysis on the pursuit task data using all motion direction conditions (the prior and outer directions). The analytical method was identical to that described above and the conclusions were similar (fig. S8, I to P).

To test whether the difference in the prior axis–projected population responses between the wide and narrow prior blocks was significant, we conducted two-sided permutation tests. We randomly shuffled the trial indices for the prior blocks and performed the TDR analysis, repeating this procedure 1000 times to obtain the null distribution of the prior axis–projected population responses. We then obtained the P value of the measure by comparing the original value with the null distribution. The prior-axis projection difference in spontaneous activity was 7.94, P = 0.001 (high contrast, monkey A); 2.23, P = 0.001 (low contrast, monkey A); 0.96, P = 0.001 (high contrast, monkey B); and 0.25, P = 0.018 (low contrast, monkey B). The prior-axis projection difference in the time window between −400 and 600 ms from the local motion onset was 6.72, P = 0.001 (high contrast, monkey A); 1.47, P = 0.001 (low contrast, monkey A); −1.18, P = 0.001 (high contrast, monkey B); and 0.62, P = 0.018 (low contrast, monkey B).

Supplementary Materials

This PDF file includes:

Figs. S1 to S11

Tables S1 and S2

View/request a protocol for this paper from Bio-protocol.