Motion Perception in the Common Marmoset

Abstract

Visual motion processing is a well-established model system for studying neural population codes in primates. The common marmoset, a small new world primate, offers unparalleled opportunities to probe these population codes in key motion processing areas, such as cortical areas MT and MST, because these areas are accessible for imaging and recording at the cortical surface. However, little is currently known about the perceptual abilities of the marmoset. Here, we introduce a paradigm for studying motion perception in the marmoset and compare their psychophysical performance with human observers. We trained two marmosets to perform a motion estimation task in which they provided an analog report of their perceived direction of motion with an eye movement to a ring that surrounded the motion stimulus. Marmosets and humans exhibited similar trade-offs in speed versus accuracy: errors were larger and reaction times were longer as the strength of the motion signal was reduced. Reverse correlation on the temporal fluctuations in motion direction revealed that both species exhibited short integration windows; however, marmosets had substantially less nondecision time than humans. Our results provide the first quantification of motion perception in the marmoset and demonstrate several advantages to using analog estimation tasks.

Keywords: decision-making; marmoset monkey; motion estimation; psychophysics; vision.

Figure 1

Visual stimuli and behavioral task. (A) Marmosets were trained to maintain fixation within a window 2° in diameter around a target presented at the center of the screen (1). A random pattern of dots was then presented within a circular aperture 7° in diameter centered on the fixation target (2). The dots moved at a speed of 15°/s in one of eight possible directions equally distributed between 0 and 360°. Coincident with the onset of the random dot pattern, eight small choice targets were presented, equally spaced around a ring, 10.6° in diameter, concentric with the central fixation target. Marmosets received a liquid reward for correctly reporting the direction of motion by making a saccade to one of the choice targets (3). On a proportion of trials, marmosets received an overt cue, consisting of a small high contrast Gabor patch, presented at the location of the correct choice target. On these trials, marmosets could obtain the reward by making an eye movement to the cued target location without integrating the motion stimulus. These cued trials served to ensure a sufficiently high rate of reward to keep the marmosets engaged with the task. Over the course of training, the proportion of cued trials was gradually reduced. (B) Sequence of trial events. After a fixation period of 200–500 ms (1), the random dot pattern appeared (2). Marmosets were required to maintain fixation on the central target for a minimum duration of 100 ms after appearance of the motion stimulus, after which the fixation target dimmed and the marmosets were free to indicate the perceived direction of motion by making an eye movement to one of the choice targets. Both the fixation point and the random dot pattern were extinguished if the marmoset broke fixation or after a maximum period of 600 ms, whichever came first (3). (C) After initial training, the number of possible motion directions was increased from 8 to 50 over the course of several weeks and the discrete choice targets were replaced by a continuous ring. The strength of the motion signal was then varied by assigning to each dot a direction drawn from a uniform generating distribution centered on the target motion direction.

Figure 2

Initial task training. (A) Total number of trials (open symbols) together with the number of completed trials (filled symbols) per session. Trial counts over 1 week after approximately 1 year (five sessions for monkey S; four sessions for monkey H) of training are shown on the right in each panel. In these sessions, the monkeys were performing the continuous version of the estimation task with 50 possible motion directions and nine possible motion SSs. (B) Initially, both marmosets based their choices on the overt cue rather than motion of the random dot pattern. However, over the course of training, the proportion of cued trials decreased as the onset of the cue was progressively delayed. (C) As a measure of performance, we computed the proportion correct—the proportion of noncued trials in each session in which the marmoset’s choice fell within the reward window (see Material and Methods). Proportion correct for both marmosets improved over the course of training. Solid lines show least-squares fits of a single exponential function.

Figure 3

Choices reflect stimulus motion direction after training. (A) Distributions of saccade end-points for noncued trials for both marmosets as kernel density plots. At each spatial location, target motion direction is represented by the hue (see inset) while the density of saccade end-points (i.e., trials) is represented by the saturation. For the trials shown, random dot patterns moved coherently in one of 50 possible directions between 0 and 360°. (B) Mean choice direction (symbols) for noncued trials plotted against target motion direction. Behavioral choices of both marmosets were highly correlated with the target motion direction. Error bars show ±1 standard deviation. (C) Distributions of angular error, the difference between the marmoset’s choice and the true motion direction on each trial, for both marmosets.

Figure 4

Motion estimation performance improved with training. (A) To quantify behavioral performance, both during training and subsequently on the main task, we modeled the distributions of angular errors as a mixture of two probability distributions: a uniform distribution (reflecting nonperceptual errors or “lapses”) and a wrapped normal distribution (reflecting errors in perceptual processing of the motion stimulus). The relative contribution of these two distributions is determined by the lapse rate, λ. Task performance is quantified by the standard deviation, σ, of the wrapped normal distribution. (B) Standard deviation of both marmosets’ errors plotted as a function of training session number. (C) Lapse rate plotted as a function of training session number for both marmosets. The lapse rate of both marmosets decreased with training. Note that we were unable to fit the mixture model in some sessions, particularly early in training when the marmosets performed relatively few noncued trials in each session. Sessions containing too few (<30) noncued trials are indicated by open symbols at the top of the axes in B and C. Symbols on the far right of each panel in B and C show standard deviation and lapse rate, respectively, after more than 1 year of training (see Fig. 2). Solid curves show least-squares fits of a single exponential function.

Figure 5

Perceptual errors vary systematically with motion SS. (A) Distributions of angular errors—the difference between the marmoset’s choice and the true motion direction on each trial—for a range of motion SSs (see Material and Methods). Signal strength, SS = 1, corresponds to coherent motion while SS = 0 corresponds to random, incoherent motion. Error distributions (bars) of both monkeys became broader as SS was reduced. The data include 17 021 trials, across all conditions, from 85 sessions for monkey S and 17 937 trials from 77 sessions for monkey H. Solid curves show the probability density defined by the mixture model (see Materials and Methods) fitted to the error distributions. (B) Proportion of correct (i.e., rewarded) trials as a function of SS. Proportion correct for both marmosets decreased as SS was reduced. In the absence of any coherent motion signal (SS = 0), both marmosets performed at the chance level (dashed line). (C) Standard deviation of the mixture model, fitted to each marmoset’s errors, as a function of stimulus strength. The standard deviation of both marmosets’ errors increased as SS was reduced. For comparison, B and C show comparable metrics for four human observers performing the same motion estimation task (see Material and methods). The human observers performed better than the marmosets over all SSs. In B and C, error bars show bootstrap estimates of the 95% CI for the corresponding metric. Solid curves in B show maximum likelihood fits of a logistic function.

Figure 6

Reaction time varies systematically with motion SS. (A) Distributions of reaction time—the time interval from onset of the motion stimulus until the marmosets’ indicated their choice—for a range of motion SSs. Arrow heads indicate the median reaction time for each distribution. (B) Median reaction time as a function of SS. For comparison, B also shows median reaction times for four human observers performing the same motion estimation task. Both humans and marmosets exhibit a typical increase in reaction time as SS was reduced. However, the increase in reaction time of the human observers was less dramatic than that of the marmosets. In B, error bars show bootstrap estimates of 95% CIs. Solid curves show least-squares fits of a hyperbolic tangent function.

Figure 7

Choices reflect recent stimulus history. (A) To assess the influence of different stimulus epochs (frames) on each subject’s choices, we estimated their temporal integration weights (temporal kernel) using psychophysical reverse correlation. For each trial, k, we computed the difference between the mean motion direction over all dots, , for each stimulus frame, j, and the target motion direction, . For each frame, j = 1, …, N, we constructed a vector containing these differences for all trials. We then computed the correlation between this vector, for each frame, with the vector containing the difference between the subject’s choices, , and the corresponding target motion directions, , to reveal the subject’s “temporal kernel”. (B) Temporal kernels for each subject for trials aligned with the onset of the motion stimulus. (C) Temporal kernels for each subject as in B, after realigning each trial with the onset of the saccade indicating the subject’s choice. Arrow heads indicate the estimated saccade dead time for each subject. For comparison, B and C also show temporal kernels for four human observers performing the same motion estimation task. Shaded regions indicate bootstrap estimates of 95% CIs. (D) Average saccade-aligned temporal kernels for marmosets and humans, accounting for differences in saccade dead time and normalizing to the peak amplitude.

Figure 8

Choices are independent of small drifts in eye position during motion presentation. (A) Horizontal and vertical eye position from a representative trial from one monkey (Monkey H, SS = 0) plotted in space (upper panel) and over time (relative to motion stimulus onset; lower panel). We often observed small drifts in eye position during presentation of the motion stimulus (green symbols in the upper panel and green shaded epoch in the lower position-vs.-time panel). (B) Eye speed versus time relative to motion stimulus onset, projected onto the motion stimulus direction and averaged over all trial for each SS. On average, the eyes remain stationary for approximately 75 ms after motion onset (longer, ~125 ms, in human observers) before drifting slowly until onset of the saccade indicating the subject’s choice. Drift speed decreased systematically as motion SS was reduced. (C) Magnitude of the drift in the direction of the stimulus motion. (D) Proportion of correct trials as a function of SS, based on the drift vector direction. Proportion correct for both marmosets and humans decreased as SS was reduced. To aid comparison, dashed lines show proportion correct for each subject based on their choices (reproduced from Fig. 5B). (E) Density plots of saccade error versus drift error to assess the extent to which systematic drift in eye position could account for the subject’s subsequent choice. For both marmosets and humans, errors in the subject’s choices were independent of errors in drift direction. Same conventions as in Figs 5–7.