Familiarity with visual stimuli boosts recency bias in macaques

Subjects

Data were obtained from two adult male rhesus monkeys (7–10 kg). Animals were housed in standard cages in the Washington Primate Research Center, with pair housing when possible (Liu, Murray & Jagadeesh, 2009). Monkeys B. and S. were trained, acquired, and raised under the same protocol (# 3275-01) as monkeys L. and G. (Akrami et al., 2008; Liu, Murray & Jagadeesh, 2009). Prior to this study, the animals were well trained in versions of the 2-alternative forced choice design used in this experiment (Liu & Jagadeesh, 2008). During each session, NHPs were positioned in a booth, and performed a two alternative forced choice-delayed match to sample task (2AFC). Each trial consisted of a series of events, displayed on a computer monitor, and ended with a response in the form of a saccade (see Experimental Design). The stimuli were displayed using CORTEX, a program for neural data collection and analysis developed at the NIH (Bethesda, MD). Both NHPs were implanted with a head restraint prosthesis, used to limit movement of the head during data collection, and a scleral eye coil to monitor eye position (DNIm Newarkm DE, USA; (Judge, Richmond & Chu, 1980)). To induce motivation for the cognitive task, freely available water was limited the day before the experiment, and fluids (water or juice drops) used as a reward following a correct trial. All animal handling, care and surgical procedures were performed in accordance with guidelines established by the NIH, and approved by the institutional Animal Care and Use Committee (IACUC) at the University of Washington (Protocol number 3275-01).

Stimuli and experimental design

For each session, one pair of photos was selected from a picture database containing images of everyday objects such as faces, plants, animals, landscapes and buildings (see Fig. 1A for some examples). All visual stimuli used for that session were then derived from those images. All images were 90 × 90 pixels, drawn from a variety of sources, such as databases and personal photo libraries. Stimuli were presented on a computer monitor with 800 × 600 resolution and a refresh rate of 100 Hz. At the viewing distance used, images subtended 4° of visual angle. The visual stimuli were either familiar to the subject (used during a previous session) or novel (never used before).

The sequence for each trial was the same (see Fig. 1C): after the monkey foveated a fixation point for a variable time between 250 and 500 ms within an invisible fixation window that was 4 degrees in diameter and centered around the fixation point, a visual cue was presented. The cue was displayed for 500 ms, and either static or dynamic (see below), and its location and size the same as for the invisible fixation window. After the cue disappeared, the fixation point was still turned on for a variable delay time between 700 and 1,200 ms, and after the choice pair was displayed, for another variable period of 400–700 ms. Only then was the fixation point turned off, signaling the NHP that it was allowed to make a choice. The NHP could express its preference by making a saccade towards either image A or image B (together forming the choice pair). The monkey received a reward for selecting the image (either A or B), that best matched the cue.

To avoid spatial bias, the order of the choice pair, located 5 degrees to the left of the fixation point was randomized: Image A was 5 degrees either up or down, and image B respectively 5 degrees down or up with respect to the fixation point (see Fig. 1C).

For trials where the cue was static (either image A or image B) reward followed the selection of the corresponding image from the choice pair (delayed match-to-sample; see Fig. 1C). For some trials, however, a dynamic cue was used: a movie snippet of 5 frames, drawn from an 11-frame long movie that was generated by morphing image A (frame 1) slowly into image B (frame 11; see Fig. 1B, blue and pink boxes). The images were morphed by using MorphX (http://www.norrkross.com/software/morphx/morphx.php), a program for morphing between two photographic images. The dynamic cues, consisting of either the first or last five frames of the larger movie, were displayed for 500 ms (100 ms per frame) and presented either in forward or reverse order. Trials with dynamic cues were rewarded when the NHP selected the target (from the choice pair) that was also presented in the cue. More specifically, trials where the cue was a movie snippet consisting of frames 1-2-3-4-5 or 5-4-3-2-1, led to reward when the animal selected image A (equal to frame 1), whereas trials where the cue was a movie snippet consisting of frames 7-8-9-10-11 or 11-10-9-8-7 led to reward when the NHP selected image B (equal to frame 11).

For a subset of the trials, the dynamic cue was ambiguous, consisting out of the five frames drawn from the middle of the 11-frame movie, thus containing neither image A nor image B (see Fig. 1B, green box). The ambiguous cue was presented either in forward (frames 4-5-6-7-8) or reverse direction (frames 8-7-6-5-4). For those ambiguous trials, where there was no correct answer, the animals were rewarded randomly at a 50% chance rate upon completing the trial.

For some of the sessions, the dynamic cue was “weighted”. We introduced this condition after we noticed—based upon preliminary results—that the NHPs were more likely to extract information from the end of a dynamic cue than from the beginning (see results). We were interested in whether weighting the movie towards the beginning, by removing the penultimate frame (frames with a red crossed box, in the example shown in Fig. 1B), would change the behavior of the NHPs. The intervention, resulting in a 100 ms long gap during the display of the cue, however made perceptually almost no difference, and consequently did not affect the behavior of the animals (as reported in the result section).

Sessions were organized in blocks of 60 trials. The first block contained only unambiguous static cues (see Fig. 1B, row 2 for an example), which was meant to motivate the subject and to make it familiar with the stimulus pair before confronting it with trails where the cues were dynamic. Subsequent blocks, also 60 trials long, consisted of all 10 conditions, randomly mixed. Those are the two static conditions (Fig. 1B, row 2), the four unambiguous dynamic conditions (Fig. 1B, movie snippets in blue and pink boxes), two ambiguous dynamic conditions (Fig. 1B, green box) always followed by reward, and the same two dynamic conditions, but never followed by reward. We collected minimum five blocks of data for each session, yielding at least 96 trials where ambiguous cues were used. We collected data from 141 sessions for monkey B. and 269 sessions for monkey S., totaling more than 200,000 trials.

Data analysis

All analyses, including averages, error estimates, and statistical significance were calculated using Matlab, version R2015a; Figures were generated using Matlab, version R2015a, and Adobe Illustrator CS6. Averages are given as unweighted mean ± SE.

To quantify the results obtained for this study, we used two equations.

The first equation is used to calculate the Preference Index (PI), a metric that quantifies temporal biases which we termed recency and primacy (see results). PI is calculated as follows: (1)${\displaystyle PI=\left(\frac{R-P}{R+P}\right)}$ where R and P represent the number of ambiguous trials within a given session where the NHP respectively selected the image most resembling the end (Recency) or beginning (Primacy) of the dynamic cue (see Experimental design and Fig. 2, green box).

A second equation was used to quantify the behavioral shift (Δ_PI), between two sessions (S1 and S2) where the same set of visual stimuli were used. We denote S1 as the session where a given set of stimuli was used for the first time and S2 as the session where the same set was used again, thus for a second time. The Δ_PI for a pair of S1 and S2 is obtained as follow: (2)${\displaystyle {\Delta }_{PI}=P{I}_{S2}-P{I}_{S2}}$ where PI_S1 and P_IS2 are the PI values (see Eq. (1)) calculated for respectively S1 and S2.

Discrimination between two visual stimuli, using static and dynamic cues

Before acquisition of the behavior data used for this study, two NHPs were first trained in the 2AFC task using only static cues (Delayed matching-to-sample). Once the NHPs routinely performed at high levels of accuracy (>90% correct), we also included 5-frame long dynamic cues (see Materials and Methods), drawn from a larger 11-frame long movie generated by morphing image A (frame 1) stepwise into image B (frame 11).

Once the NHPs also mastered trials with dynamic cues (>90% correct), we switched from training to working mode, and composed a complete new set of image pairs, from which we selected a single pair for each session (see Fig. 1A for some example pairs).

The animals performed well on trials with dynamic cues, whether the target frame was included at the beginning (Figs. 2A and 2B, purple bars; forward direction) or end (green bars; reverse direction) of the dynamic cue, as illustrated by the success rates: 86.0 ± 1.1% and 86.3 ± 1.0 (monkey B); and 93.6 ± 0.7 and 94.1 ±0.6% (monkey S) for respectively forward and reverse playing directions. A two sample t-test comparing the means of the success rates computed for each playing direction revealed that the playing direction of the cue had no effect on performance (monkey B.; p = 0.80; N = 107 and monkey S; p = 0.40; N = 131). As expected, the NHPs were not able to figure out a strategy allowing them to receive more reward than predicted by chance (Figs. 2A and 2B, gray bars) for trials with ambiguous cues, demonstrating that the reward was truly random: 49.5 ± 0.5% for monkey B., and 49.9 ± 0.3% for monkey S. To examine whether performance changed over the course of a session, which might signal fatigue, we also computed the proportion of rewarded trials as a function of trial number for a given condition. To that extent, we normalized each session with respect to the total number of trials. The results, based upon an average of all sessions (Figs. 2C and 2D) suggest that the primates did neither change behavior nor strategy with respect to any of the conditions during the course of a session.

Not displaying the penultimate frame (weighted cues; see Materials and Methods) caused a small interruption in the sequence, and potentially a perceptual shift towards the beginning of the dynamic cue (see Materials and Methods). The intervention, however, did affect neither behavior nor performance as illustrated by a two sample t-test comparing the means of the success rates for sessions of each condition: there was no statistical difference between sessions using weighted versus unweighted movies for the forward direction (p = 0.06 for monkey B. and p = 0.14 for monkey S.), nor for the reverse direction (p = 0.21 for monkey B. and p = 0.76 for monkey S.).

Recency and primacy effects

Because the performance of the NHPs for trials with informative cues was very satisfactory, not sensitive to the direction of the dynamic cue, and consistent over the course of a session (see Fig. 2), we were confident that the NHPs, even when facing ambiguous cues, would show non-random behavior by continuing to apply the learned rules for making a “correct“ response. To probe the behavior with respect to those “impossible” trials, the ultimate aim of this study, we will from here onwards only consider the data from those trials. Moreover, for this subset of the data, we will no longer focus on correct versus non-correct identification of the target which was determined random and therefore not informative, but instead on the interpretation of the cue by the animal. Would it rely, trying to identify the “correct” target, on the information extracted from the beginning (more similar to one image from the choice pair), or rather from the end (more similar to the other image) of the dynamic cue? To quantify the weight of each of those effects, we calculated a Preference Index (PI; see Eq. (1) in Materials and Methods). This metric yields a value between −1 and +1 for each session. A value smaller than 0 reveals that for this session the primate was more likely to engage in primacy bias, while a value greater than 0 indicates that the recency effect prevailed. It should be noticed that the terms “primacy” and “recency”, for this study, are not used in exactly the same way as for serial position effect studies (Murdock & Bennet, 1962), where they refer to the tendency of an individual to respectively best recall the first and last items of a list.

Because weighting of the stimuli had no significant effect on the PI values, we merged both conditions for further analysis. Computation of the mean PI value, averaged across all 141 sessions from monkey B. and all 269 sessions from monkey S. revealed that both primates engaged in behavior that slightly, but consistently, tilted towards recency bias (Figs. 3A and 3B). The averaged PI computed for monkey B. was 0.08 ± 0.01 (inset Fig. 3A) and 0.09 ± 0.01 for monkey S. (inset Fig. 3B). Although the effect size appears small, it was statistically very significant for each animal, as demonstrated by using a non-parametric sign test yielding p-values of 1.7 ×10⁻⁶ and 3.2 ×10⁻⁹ for monkeys B. and S. respectively. This tendency towards recency is also illustrated by the distribution of the PI values (Figs. 3A and 3B), where the bin with value 0 (marked by a vertical black line) represents sessions where the recency and primacy effects were equally balanced.

It is possible that towards the end of the session NHPs, with respect to ambiguous trials, would try “less hard” to find the “correct answer”. If so, then the PI value would converge towards 0 over the course of a session, consistent with random and effortless response selection. Taking the same approach as we did to study behavior with respect to unambiguous trials over the course of a session (Figs. 2C and 2D), we also computed the PI value as a function of normalized trial number. The results (Figs. 3C and 3D) suggest that also in this case, the behavior of the NHPs did not change across the duration of a session.

Taken together, the results shown in Fig. 3 suggest that the NHPs, despite receiving reward at random, did not choose randomly but used a choice strategy as illustrated by a small but very significant recency effect.

The effect of stimulus familiarity on behavior bias

To study the effect of familiarity of the visual stimuli, we examined whether behavior exhibited during a session where a given pair of images was novel (denoted S1) shifted when those same stimuli were used again, later, during a second session (denoted S2). Across both NHPs, we identified 163 instances where the same stimuli were used for two different sessions by the same animal (2 ×49 sessions from monkey B. and 2 × 114 sessions from monkey S.). The averaged PI value of all S1 sessions (Figs. 4A and 4B) was positive for both animals (0.008 ± 0.024 and 0.062 ± 0.015 for monkeys B. and S. respectively), consistent with recency bias. This was also the case for all S2 sessions (0.130 ± 0.023 and 0.099 ± 0.017 for monkeys B. and S. respectively). A paired t-test between the PI values of S1 and S2 from each pair, however, revealed that this recency bias strengthened significantly with stimulus familiarity, for both monkey B. (p = 0.0001), and monkey S. (p = 0.041). This effect of familiarity is further visualized in Fig. 4B, where each of 163 data points (green for monkey B and light purple for monkey S.) indicate a pair of sessions (S1 and S2) where the same stimuli were used. The coordinates of each data point correspond with the PI values of S1 and S2.

Those results are counterintuitive, because it was expected that the NHPs, after gaining experience with the stimuli, would begin to unmask the dubious nature of the ambiguous trials, and start selecting the “target” more randomly, resulting in a PI value closer to 0 (the value associated with chance). Instead, reusing the same stimuli strengthened the recency effect, suggesting that the primates, instinctively, increased the use of a strategy that has not proven to pay off.

This observation gives rise to a new question: “is the lag time between a pair of sessions (S1 and S2) critical for the observed phenomenon?”

To answer that question, we computed Δ_PI (see Materials and Methods for details), between any pair of sessions where a given set of stimuli was used for respectively the first (S1) and second time (S2). This yielded a single value for each of the 163 pair of sessions revealing whether familiarity with a given set of stimuli would not affect behavior (values close to 0), would induce a shift towards primacy bias (negative value), or cause a shift towards recency bias (positive value).

To take the elapsed time (measured in days) between S1 and S2 into account, we divided the 163 pairs of sessions into 3 bins (Fig. 4D). The first bin contained all the pairs of sessions where S1 and S2 were separated by 3 days or less (12 pairs for monkey B; 48 pairs for monkey S.); the second bin all pairs of sessions separated by 4 to 6 days (18 pairs for monkey B. and 28 pairs for monkey S.); and a third bin for sessions separated by more than 6 days (19 pairs for monkey B. and 38 pairs for monkey S.). Interestingly, a paired t-test between the S1 and S2 sessions revealed that the behavioral shift is not statistically significant (for neither NHP) for the pairs grouped in the first bin, but significant (monkey S.; p = 0.030) or even extremely significant (monkey B.; p = 0.0003) for the pairs that make up the second bin. The effect then seems to fade away again for sessions separated for more than 6 days (3rd bin), still significant in monkey B. (p = 0.034) but not in monkey S. (p = 0.313).

If the primates over the course of many sessions, recorded over a period of 10 months, exhibited increasingly more recency behavior, then the observed behavioral shift (Figs. 4A–4C) could simply be explained due to the fact that S2 was recorded at a later date than S1, and not because of familiarity with the stimuli. Plots of PI versus recording day (Figs. 5A and 5B), however, show that the animals did not change their behavior over the course of ∼300 days of data collection. The slope of the linear regressions (black lines in Figs. 5A and 5B) are not statistically different from zero (p = 0.09 for monkey B., and p = 0.96 for monkey S.).

Because those unambiguous trials were rewarded at random, it is not impossible that for some sessions, a given strategy (for instance recency) was rewarded more often, or less often, than what would be expected by chance. To investigate whether our results could be explained by a reinforcement history, we plotted the behavior shift between S1 and S2 (ΔPI) as a function of how often a given strategy, primacy (Fig. 5C) or recency (Fig. 5D), was rewarded during performance of S1. The insets of each figure shows how the linear regression of the data would look if there was a strong effect of reinforcement history. For instance, always (never) receiving reward for choosing recency during S1 would trigger the animal to rely more (less) on recency during S2, which would then result in a more positive (negative) ΔPI (inset Figs. 5D). The analysis (5C and 5D), however, demonstrates that our results cannot be explained by reinforcement history: the slopes of the linear regressions (green line for monkey B. and purple line for monkey S.) in Fig. 5C are not statistically different from zero (p = 0.55 for monkey B., and p = 0.27 for monkey S.); this is also the case for the regressions shown in Fig. 5D (p = 0.19 for monkey B., and p = 0.92 for monkey S.)

Taken together, the results, shown in Figs. 4 and 5, suggest that the cognitive processes that took place while the visual stimuli were used for the first time, altered behavior during a later session when those same stimuli were used again. Interestingly, the influence appears to be mild when familiarity with the stimuli is only established recently, but strong when stimuli used for the first time are used again after 4 to 6 days. Unsurprisingly, the effect decreased with longer time between both sessions.