Capuchin monkeys (sometimes) go when they know: Confidence movements in Sapajus apella

Abstract

To test for evidence of metacognition in capuchin monkeys (Sapajus apella), we analyzed confidence movements using a paradigm adapted from research with chimpanzees. Capuchin monkeys provide an interesting model species for the comparative assessment of metacognition as they show limited evidence of such cognitive-monitoring processes in a variety of metacognition paradigms. Here, monkeys were presented with a computerized delayed matching-to-sample (DMTS) memory test in one location but were rewarded for correct responses in a separate location. Movements could be made from one location to the other at any time, but movements between a response and reward feedback may reflect confidence in the accuracy of the response. Critically, DMTS tests included occasional “no sample” trials where monkeys' performance was at chance when the trial started without a sample and a 1-s interval to the response options. We predicted that monkeys would (1) perform less accurately (and less confidently) at longer retention intervals, (2) move to the dispenser early more often on trials completed correctly than incorrectly, and (3) show a relation between faster response latency and early movements. Analyses of response times and “go” or “no go” confidence movements before feedback to the reward location suggested that the monkeys were capable of monitoring confidence in their responses. However, their confidence movements were less precise and less flexible than chimpanzees. Overall, this paradigm can reveal potential metacognitive abilities in nonhuman animals that otherwise demonstrate these abilities inconsistently.

Introduction

For more than two decades, comparative psychologists and others studying cognitive processes in animals have focused on a particular aspect of cognition that is not about primary perceptual mechanisms, conceptual classifications, or memory processes, but instead is focused on the possibility that animals are aware of their knowledge states and can adjust behavior on the basis of this awareness. This is called metacognition (Dunlosky & Bjork, 2008; Flavell, 1979; Metcalfe & Shimamura, 1994; Nelson, 1992; Nelson & Narens, 1990; Schwartz, 2009). Metacognition is defined in various ways, sometimes muddying the waters of interpretation of animals' behavior in so-called metacognitive tasks. However, a growing literature suggests that, at minimum, some species (i.e., rhesus monkeys, orangutans, gorillas, rats) do learn to avoid difficult perceptual classification trials (perhaps because they are aware of the difficulty they face) by escaping those trials (e.g., Smith et al., 1995; Smith, Beran, Redford, & Washburn, 2006; Smith, Coutinho, Church, & Beran, 2013; Smith, Redford, Beran, & Washburn, 2010; Smith, Shields, Schull, & Washburn, 1997; Suda-King, 2008; Suda-King, Bania, Stromberg, & Subiaul, 2013; Yuki & Okanoya, 2017). Some species (i.e., rhesus monkeys, baboons, pigeons, rats) re-study information they are likely to have forgotten prior to an impending memory test or avoid taking the test altogether for exactly those stimuli that are most difficult to remember (e.g., Adams & Santi, 2011; Basile, Schroeder, Brown, Templer, & Hampton, 2015; Brown, Basile, Templer, & Hampton, 2019; Brown, Templer, & Hampton, 2017; Foote & Crystal, 2012; Hampton, 2001; Malassis, Gheusi, & Fagot, 2015; Templer & Hampton, 2012; Templer, Lee, & Preston, 2017). Some species (i.e., rhesus monkeys, chimpanzees, orangutans, dogs, rats, capuchin monkeys) seek additional information before choosing a response option when they cannot possibly know the correct response without that information (e.g., Belger & Bräuer, 2018; Beran & Smith, 2011; Beran, Smith, & Perdue, 2013; Call, 2010; Call & Carpenter, 2001; Hampton, Zivin, & Murray, 2004; Kirk, McMillan, & Roberts, 2014; Marsh and MacDonald, 2012a, Marsh and MacDonald, 2012b; Vining & Marsh, 2015). Furthermore, some species (i.e., rhesus monkeys) show that they can provide confidence-like ratings to their own responses to tests that align with objective performance levels, including those that are made prospectively or retrospectively (e.g., Morgan, Kornell, Kornblum, & Terrace, 2014; Shields, Smith, Guttmannova, & Washburn, 2005). And, some species (i.e., rhesus monkeys, western scrub-jays) calibrate their escape responses to avoid having to complete difficult trials or their hint-seeking behaviors based on the rewards at risk and their likelihood of responding correctly (e.g., Kornell, Son, & Terrace, 2007; Watanabe & Clayton, 2016; Zakrzewski, Perdue, Beran, Church, & Smith, 2014). This growing literature continues to generate strong debate about the proper interpretation of these kinds of behaviors (Beran, Brandl, Perner, & Proust, 2012; Carruthers, 2008, Carruthers, 2009; Comstock & Bauer, 2018; Crystal, 2014; Crystal & Foote, 2009; Hampton, 2009; Kornell, 2009, Kornell, 2014; Le Pelley, 2012; Smith, 2009; Smith, Couchman, & Beran, 2012; Smith, Zakrzewski, & Church, 2016), but there is consensus that the empirical database continues to grow in terms of positive reports of success in these kinds of tasks.

Some of the strongest evidence has emerged for metacognitive monitoring within nonhuman primates, as evidenced by the many citations to papers with the great apes and monkeys in the previous paragraph. However, the breadth of assessment within the order Primates remains fairly limited, with only two of four great apes (chimpanzees and orangutans), two Old World monkey species (rhesus macaques and baboons), and one New World monkey species (capuchin monkeys) having been tested. Despite the limited variability among species tested thus far, there is an intriguing pattern of results that has emerged – capuchin monkeys only sometimes perform consistently with the great apes and Old World monkeys. The earliest assessments with capuchin monkeys focused on information-seeking tasks (Basile, Hampton, Suomi, & Murray, 2009; Paukner, Anderson, & Fujita, 2006; Vining & Marsh, 2015), in which capuchin monkeys sometimes searched an opaque container for food when they should (e.g., when they did not see the food baited) but also often searched when they should not have needed to do so (e.g., the container was transparent). In memory monitoring tasks, capuchin monkeys also failed to show the metacognitive-like patterns of responding shown by rhesus macaques (Fujita, 2009). Beran, Smith, Coutinho, Couchman, and Boomer (2009) presented capuchin monkeys with an absolute classification task and trained monkeys to classify stimuli as a function of the degree of pixilation that made up those stimuli (sparse vs. dense). For the metacognitive response, monkeys were presented with the option to skip to the next trial free of punishment (which was a time-out) but also at the cost of losing any positive reinforcement (food reward). Unlike rhesus monkeys and humans (e.g., Smith et al., 1997), capuchin monkeys did not adopt the metacognitive response, failing to selectively escape difficult trials.

The previous psychophysical task in which monkeys were required to classify stimuli as ‘sparse’ or ‘dense’ (Beran et al., 2009) was then used in a series of additional studies to determine what might account for the seeming “divide” among the primate species tested. First, it was demonstrated that capuchin monkeys may be more inclined to make primary responses (those reinforced or punished based on correctness of the response) when the number of choice options was small, and thus the chance of succeeding was higher by chance alone. When tasks were adjusted to make guessing less profitable, capuchin monkeys then (sometimes) added escape responses to their behavioral repertoire and did so for exactly those trials that they should escape if difficulty was their guide (Beran, Perdue, Church, & Smith, 2016; Beran, Perdue, & Smith, 2014). This suggested that when there were only two stimuli to choose between (and guessing was 50% correct), the capuchin monkeys may have preferred a risky guess strategy over the safer trial-escape strategy, especially when the escape strategy did not benefit them with positive reinforcement. This may not reflect a lack of metacognitive capacity but rather a higher threshold for engaging in such monitoring. However, removing reinforcement from the initial choice to respond or escape and reserving it for a subsequent unrelated task still did not evoke uncertainty responses to difficult or impossible to perform trials in this species (Perdue, Church, Smith, & Beran, 2015), and this did suggest a real limitation for engaging in metacognition in this species.

Given the variable nature of the existing data, we are left at present with the curious case of the capuchin monkey (Smith, Beran, Couchman, Coutinho, & Boomer, 2009; Smith, Smith, & Beran, 2018), and why either (1) they are not capable of metacognitive responding, or (2) they are reluctant demonstrators of a psychological metacognitive state that they might experience. This reluctance may seem confusing given that some non-primate species have demonstrated strong metacognitive-like patterns. For example, evidence from rats shows a variety of sophisticated behavioral patterns in tasks designed to assess metacognition (e.g., Crystal & Foote, 2009; Foote and Crystal, 2007, Foote and Crystal, 2012; Kirk et al., 2014; Templer et al., 2017; Yuki & Okanoya, 2017). However, the comparative literature also includes at least one other species that could be classified as either a non-metacognitive species or a reluctant metacognitive responder. Pigeons show the same inconsistency and variability of performance that capuchin monkeys have shown, with positive reports (Adams & Santi, 2011; Castro & Wasserman, 2013; Nakamura, Watanabe, Betsuyaku, & Fujita, 2011) matched by partial or full failures (Inman & Shettleworth, 1999; Iwasaki, Watanabe, & Fujita, 2013; Roberts et al., 2009; Sutton & Shettleworth, 2008) and debate about the proper interpretation of the empirical data (e.g., Smith, 2009; Sole, Shettleworth, & Bennett, 2003; Zentall & Stagner, 2010). And, it is important to note that we have barely begun to assess the broadness or depth of animal metacognition, with such restricted numbers of primate and non-primate species tested.¹

Recently, we developed a new measure of metacognition that involves the use of so-called confidence movements. The original task was presented to chimpanzees (Beran et al., 2015), and it involved requiring the subjects to work on a primary task in one location and receive a reward (if earned) in another location. The basic procedure is to have the subject complete a computer task of varying difficulty across trials (e.g., a memory test with variable retention delays) and, after a response is made, to insert a delay before any feedback is given by the computer. During that delay, the subject can move (or not) towards a distant location where reward is dispensed. After the delay, feedback is given about the outcome of the trial, and then after another short delay, a food reward is dispensed (if earned). The key manipulation in this approach is that the reward is forfeited if not obtained at the moment it is dispensed. In other words, if the subject has not previously moved to the reward site before it is delivered, the reward is lost. One must be at the dispenser in time, and if one waits for feedback from the computer, the time left to travel to the dispenser makes it effortful and somewhat risky that the reward might be lost. Moving early to the reward dispenser, during the delay before feedback, affords plenty of time to be in place for a reward, if one ultimately is dispensed. Thus, if one has high confidence in being correct, moving early is optimal. Low confidence might evoke early moves also, but at the potential cost of wasted movement if no reward is delivered because the response was incorrect. Failures to move early to the reward dispenser when correct mean that the subject must move very quickly and effortfully to still obtain the reward, but if the trial is incorrect, the next trial can be started immediately at little or no movement cost. Thus, the optimal pattern is to go early to the reward dispenser when confident about the outcome of the trial, and this should generate a pattern of more early moves on correct trials than on incorrect trials. Chimpanzees show exactly this pattern, across a series of tests (Beran et al., 2015). To date, they are the only nonhuman species to be given this task.

There is reason to predict that capuchin monkeys might show a pattern similar to that of chimpanzees, even if in general they have a tendency to want to respond rather than escape or avoid primary responses in metacognitive tasks. In the present task, subjects must always make the primary response to the computer task (completion of the memory trial). However, the metacognitive response lies in the behavior that follows the completion of the trial – movement (or not) towards the food dispenser. It is true that one might predict overconfidence being aligned with a higher level of risk tolerance (i.e., monkeys move towards the reward dispenser prior to computer feedback regardless of performance), but even in that case, monkeys might come to learn that early movements are differentially reinforced on the basis of primary task performance. And, when confidence is low, remaining at the computer would allow for quicker re-engagement of the computer task. Thus, we modified the task of Beran et al. (2015) so that it could be used with capuchin monkeys, in an effort to determine whether they also might show a pattern of early movements that matched objective performance. We also anticipated an outcome more aligned with a metacognitive pattern given that this task uses a more ecologically relevant response of movement, rather than a more artificial response that is constrained to the computer task itself. Note that we are not arguing that the current approach is ideally suited to capuchin monkeys alone. The work with chimpanzees indicates it is likely valuable with any species that moves towards food, and that makes judgements about likelihood of food availability. The point here is that capuchin monkeys are, at best, “reluctant” to engage in metacognitive-like patterns of responding in previously-used tasks, and so this task may provide more “scaffolding” for metacognitive patterns to emerge. And, this task engages natural responses such as moving to food as a function of whether there is the expectation of food in that location in order to gauge potential metacognitive abilities.

Monkeys completed a delayed matching-to-sample (DMTS) memory test with variable retention delays and were rewarded at a distant location away from the computer. For the MTS trials, we recorded trial outcome (correct/incorrect) and latency to make a response as a function of retention interval (1, 2, 4, 7, and 10-s). We recorded their movement patterns throughout trials, to measure not only confidence but also their attentiveness to trials in the primary memory task. We also presented monkeys with a crucial trial type – the no-sample trial. In tests of metamemory in nonhuman animals, the no-sample trial is important to include because it offers a short delay duration but no possible way for the animal to recognize or know the matching stimulus. For example, subjects are not provided with a sample image prior to presentation of the match options in this trial type; thus, they must guess which match option is the correct answer. This test controls for the possibility that a subject learned that shorter retention intervals lead to reinforcement. A subject that has learned that short intervals yield a high rate of reinforcement may move to the dispenser erroneously following a no-sample trial rather than moving to the dispenser as a function of memory strength for the stimulus (see Hampton, 2001, for a discussion of no-sample trials in metacognitive testing). However, a metacognitive subject that instead is aware of their performance would not expect to be reinforced on these no-sample trials. Subsequently, performance on no-sample trials should show low confidence in terms of limited or no movement to the reward dispenser.

We predicted that capuchin monkeys would show a systematic decrease in accuracy across the increasing retention intervals on the DMTS memory test, aligned closely with the varying delay intervals (i.e., the longer the delay, the poorer the performance). We predicted a greater likelihood of making a confidence movement on correctly completed trials, but we also predicted a relationship between latency in making the memory task response and confidence movements. Specifically, if monkeys remembered a sample or recognized one stimulus as more familiar,² they should respond more quickly to the memory test than if they did not, and those faster responses to complete the memory test should be predictive of confidence movement on those trials. This also could show an interesting outcome. Some trials could be correctly completed, but because of monkeys “getting lucky” after randomly responding (and perhaps taking longer to do that). In those cases, response latencies to the memory test should be slower compared to correctly completed trials for which monkeys had a strong memory trace or sense of familiarity for the sample. This outcome would show that response latency was dissociated from performance but was predictive of confidence movements in a way that reflects confidence in some correctly completed trials more than in other correctly completed trials.