What does preferential viewing tell us about the neurobiology of recognition memory?

Abstract

The two tests most widely used in nonhuman primates to assess the neurobiology of recognition memory produce conflicting results. Preferential viewing tests (e.g., visual paired comparison) produce robust impairments following hippocampal lesions, whereas matching tests (e.g., delayed nonmatching-to-sample) often show complete sparing. Here, we review the data, the proposed explanations for this discrepancy, and then critically evaluate those explanations. The most likely explanation is that preferential viewing tests are not a process-pure assessment of recognition memory, but also test elements of novelty-seeking, habituation, and motivation. These confounds likely explain the conflicting results. Thus, we propose that memory researchers should prefer explicit matching tests and readers interested in the neural substrates of recognition memory should give explicit matching tests greater interpretive weight.

Main text

The neurobiology of recognition memory is still unresolved

Ask three neuroscientists which brain areas are necessary for recognition memory, and you will likely get three different answers. Recognition memory is usually operationalized as the ability to indicate whether a current stimulus occurred previously (e.g., [1]). Despite decades of research on the neural underpinnings of recognition memory, this surprisingly simple question continues to produce conflicting results. For example, evidence exists that the hippocampus is necessary (e.g., [2,3]) and that it is unnecessary (e.g., [4,5]). Similarly, the entorhinal cortex is either critical for recognition [6], contributes modestly [7], or is unnecessary [8]. The vague familiarity component of recognition memory is variably attributed to the hippocampus [9,10], the perirhinal cortex [11], the entorhinal cortex [12,13], or the amygdala [14]. Similarly, the detailed recollection component has been attributed to the hippocampus [9,15,16], the entorhinal cortex [17], or other structures [18,19].

One reason studies of recognition memory produce conflicting results is because researchers use two different paradigms: preferential viewing and matching. Preferential viewing – typically assessed with a visual paired-comparison task (VPC) – measures the time a subject looks at novel vs. familiar images; it requires no training (Figure 1a). Matching – typically assessed with either a delayed matching- or nonmatching-to-sample task (DMS/DNMS) – requires explicit choices and substantial training (Figure 1b). In theory, multiple paradigms should produce robust convergent evidence. In practice, conflicting results necessitate examination of whether the paradigms are actually measuring the same process. Here, we examine the discrepant evidence from direct comparisons of these tests in human and nonhuman primates and critically evaluate the explanations for this discrepancy. Also see Box 1 for a summary of similar evidence in rodents.

Figure 1. Two common tests of recognition memory: Visual Paired Comparison (VPC) and Delayed Nonmatching-to-Sample (DNMS).

Although the two tests are both commonly used, the last two decades have seen VPC become the dominant paradigm. a) VPC involves minimal training, does not make reward contingent on memory accuracy, and relies on subjects’ preexisting preference to look at novel items. The schematic illustrates the idea of VPC as implemented in a nonhuman primate experiment. In this trial, the monkey sees a yellow triangle during the first free-viewing period, sometimes called familiarization. After the memory delay, the monkey is shown the familiar image and a novel image. If the subject preferentially looks at the novel green heart during the second free-viewing period, sometimes called test, researchers infer that it must remember the yellow triangle. Some variants use two test periods to counterbalance the location of the novel image. Preferential viewing is interpreted as remembering and indiscriminate viewing is interpreted as forgetting. b) DNMS involves substantial training and makes rewards contingent on memory accuracy. In the trial depicted in the schematic, the monkey interacts with a green cylinder during study. After a memory delay, it is shown the familiar object and a novel object. It has learned through trial and error that displacing novel items will reveal food. Thus, correct choice of the novel item is interpreted as remembering and indiscriminate choice is interpreted as forgetting. The logic of delayed matching-to-sample (DMS) is similar, but the monkey has been trained that food at test is under the object seen at study. These are example trials and variants exist of each test.

Box 1 -. Preference and matching tests provide divergent evidence in rodents.

Although this article is focused on evidence from primates, it is relevant to note that evidence from rodents also supports the conclusion that preference tests differ from choice tests. The novel-object preference (NOP) test (sometimes also referred to as novel object recognition, NOR, or spontaneous object recognition, SOR) in rodents is analogous to the VPC test in primates, with the major difference being that rodents are also permitted to explore both stimuli via paws/whiskers/smell during the NOP test. As in primates, NOP tests have been directly compared to matching tests, in which food is contingent on selection (or avoidance) of the remembered object. For example, control rats who demonstrated both novelty preference in a NOP test and accurate choice of the novel item in a modified DNMS test showed no correlation between recognition scores on the two tests [78]. In another study, rats with perirhinal cortex lesions showed impaired recognition when measured with NOP, but intact recognition when measured with a modified DNMS test [30]. In this study, although some experimental groups did show significant novelty preference in the NOP test, the control group did not, which the authors attributed to poor validity of the NOP test itself. Lastly, infusions of estradiol into the rat perirhinal cortex improved performance on NOP tests but decreased performance on matching tests [76,77]. As in humans and monkeys, it appears that rodent novelty preference and matching paradigms are not equivalent measures of recognition memory.

Preferential viewing tests and matching tests provide divergent evidence in monkeys

For memory, one of the most commonly investigated structures is the hippocampus. In monkeys, the effects of hippocampal lesions on recognition as measured by VPC and matching tests have been directly compared three times (Figure 2a-c). These comparisons are not perfect; some aspects of testing differ between the tasks, such as the use of 3D physical objects for the matching tests and 2D photographs or line drawings for the VPC test [e.g., 3,20]. Given that some proposals highlight the importance of the stored content (e.g., objects, color patches, scenes, faces) over the storage process (e.g., explicit, implicit, working, priming) [21,22], it will be important for future studies to more closely match the content of the t0-be-remembered items. Nevertheless, the robust conclusion from the most direct comparisons available is that VPC tests reveal major recognition impairments following hippocampal damage, even at short delays, but matching tests show either minor deficits or complete sparing.

Figure 2. When directly compared within the same subjects, hippocampal damage produces profound recognition impairments when measured with VPC but usually leaves recognition spared when measured with DNMS.

Performance on VPC (left column) and DNMS (right column) in (A) monkeys with selective, excitotoxic lesions [20], (B) monkeys with radiofrequency lesions [3], (C) monkeys with aspiration lesions [21], and (D) a human amnesic patient with robustly spared explicit recognition [23]. Solid black lines represent performance of control subjects and colored dashed lines refer to subjects with brain damage. Shaded regions indicate ± SEM.

The strongest comparison of VPC and matching used monkeys with excitotoxic, fiber-sparing lesions [20]. In VPC, monkeys with hippocampal lesions were normal at retention intervals up to 30 seconds, but fell to chance at retention intervals of 60 seconds of longer (Figure 2A, left). In DNMS, monkeys with hippocampal lesions were normal at retention intervals up to 10 minutes, with list lengths of up to 10 items, or with a distractor motor task during the retention interval (Figure 2A, right). Only the combination of a 10-minute delay with the disruption of being moved to the home cage and returned to the testing chamber during the delay interval produced an impairment in DNMS, but performance still remained above chance. Thus, hippocampal lesions produced no deficit on most trials with a small deficit at the most difficult trials if measured by DNMS, but abolished memory at short delays if measured by VPC.

The second comparison used radiofrequency lesions, which are less selective than excitotoxic lesions, and produced the most similar results between the two tests [3]. In both tests, monkeys were impaired at retention intervals of one minute or more (Figure 2B). However, the lesion effect size for VPC was larger than that for DNMS (mean Cohen’s d for all impaired delays: VPC = 2.59, DNMS = 1.75) and three of the five lesioned monkeys had scores below even the lowest scoring control monkey at retention intervals of only 10 seconds in VPC. Thus, hippocampal lesions produced memory deficits in both tests, but larger memory deficits in VPC.

The third comparison used nonselective aspiration lesions [23]. In VPC, monkeys with hippocampal damage were normal at a retention interval of 10 seconds, but at chance with retention intervals of 30 seconds or longer (Figure 2C, left). In DNMS, monkeys were normal at retention intervals up to 10 minutes, the longest delay reported (Figure 2C, right). Thus, hippocampal lesions produced no deficit at any retention interval if measured by DNMS, but abolished memory at short delays if measured with VPC.

A recent meta-analysis supports this discrepancy [24]. Across 14 studies using 143 monkeys, hippocampal lesions impaired memory when recognition was tested with VPC but not when tested with DNMS/DMS. In contrast, performance in DNMS/DMS tasks was impaired by perirhinal cortex damage. The combination of this dissociation at the meta-analytic level, in which subjects and task parameters differed across studies, and the direct comparisons, in which the same subjects were tested in both paradigms, provides converging evidence that preferential viewing tasks and matching tasks are not testing recognition memory in the same way.

Preferential viewing tests and matching tests provide divergent evidence in humans

Data from humans are particularly informative because humans can declare their memories. Because matching tests usually require declarative responses, they serve as the benchmark of explicit recognition against which the validity of more implicit preferential viewing tests can be judged. The most informative comparison comes from the amnesic patient Y.R. [25]. Y.R. has an approximately 46% reduction in her hippocampal volume, persistent daily memory problems, and the allocentric spatial memory deficits predicted by hippocampal damage. Most interestingly, Y.R. shows normal recognition memory. In one striking comparison, Y.R. was impaired on 32 of 34 recall tests, but only 4 of 43 recognition tests [26]. These recognition tests spanned different formats, stimuli, and response types, demonstrating that Y.R.’s declarative recognition is robustly spared. In the VPC test, Y.R. was normal at a retention interval of one second, but her memory was at chance with retention intervals of five seconds or longer [25; Figure 2D, left]. In a matching test used for direct comparison, which was similar to tests used with monkeys in that it was a forced choice test using photographs, Y.R. was at 100% accuracy at all retention intervals (Figure 2D, right). Thus, Y.R.’s robust sparing of declarative recognition memory extended to matching tests but not VPC tests.

In control human subjects, VPC also provides different results than explicit matching tests. In a recent comparison, participants’ preferential viewing was only modestly correlated with their own explicit matching of the same images; matching accuracy explained just 27-33% of the variance in viewing preference, depending on encoding depth [27]. In addition, viewing preference only correlated with matching during a narrow window from 0.5 - 1.5 seconds within the viewing period. Thus, it is unclear how much this informs the validity of other VPC tests that typically use longer viewing periods. In another influential comparison, participants’ preferential viewing did not correlate with their own explicit matching accuracy for the same stimuli [28]. Indeed, their viewing preferences were similar for remembered and forgotten images. Thus, across control humans and humans with selective brain damage, preferential viewing is, at best, a mediocre correlate of explicit matching.

Evaluating interpretations of VPC

Researchers have interpreted preferential viewing tests in five major ways.

1. VPC tests recognition memory [e.g., 27-29]. According to this view, the logic of VPC is straightforward and so should be its interpretation. It is also sensitive to hippocampal damage, as predicted by some theories, which confirms its validity.

   In our view, there is strong basis to claim that this interpretation is incorrect. Results from VPC differ from those of other recognition tests [3,20,23-25,27,28] demonstrating that VPC is not just another recognition test. The argument that VPC is valid because it is sensitive to hippocampal damage presupposes that the hippocampus is necessary to recognize items. This reasoning is circular. Whether the hippocampus, or any structure, is essential for recognition memory should not be a metric of test validity but an empirical matter. Regardless, it is clear that VPC either tests recognition in a unique manner or tests things other than recognition.

2. VPC is a test of recognition memory that is especially sensitive to hippocampal damage [e.g., 3,20]. Hippocampal lesions affect VPC more than matching tests because VPC is the more sensitive test.

   This is true, at least in primates, but we argue that it is uninformative. Saying that VPC is more sensitive than matching is not an explanation, but rather a restatement of the results; it does not identify the reasons for the differences in the tests’ sensitivity. Notably, although VPC is more sensitive than DNMS to hippocampal damage in primates, the same is not the case in rodents (Box 1); most rodent studies of hippocampal damage show either minimally blunted or completely spared novelty preference [reviewed in 30].

3. VPC tests the detailed recollection aspect, as opposed to the vague familiarity aspect, of recognition memory [e.g., 25]. Patient YR is impaired on recall tests, which rely more heavily on recollection, and spared on recognition tests, which can be supported by familiarity. Thus, her impairment on VPC shows that VPC assesses recollection.

   We argue that this is unlikely. There is strong evidence linking the hippocampus to detailed recollection [e.g., 15,16,31-33], so it follows that VPC’s sensitivity to hippocampal damage might be because VPC tests recollection [25]. However, this explanation is at odds with the task demands. Recollection is typically described as vivid and detailed [34, but see 35]. Because VPC does not explicitly require subjects to encode stimulus information, it should encourage shallow encoding and produce memory traces that are weak or vague. For example, subjects are often told to passively view stimuli “as if watching tv” [e.g., 25]. This passive viewing empirically elicits shallow encoding and results in low measures of recollection. In the literature on levels of processing, when encoding instructions shift from encouraging deep processing to encouraging shallow processing, recollection decreases more than familiarity [34]. In a direct comparison, incidental encoding instructions produced weaker recollection than intentional encoding instructions and resulted in vague memory traces that were easily confused with perceptually similar lures [36]. In a comparison of the effects of re-studying material via either incidental exposure or intentional retrieval, incidental exposure produced lower recollection estimates later than did intentional retrieval [37]. Indeed, studies of recollection and familiarity that use only incidental encoding have reported that subjects produce notably few recollection responses and the authors of these studies have speculated that this low level of recollection is due to the incidental encoding instructions [38]. Thus, although VPC likely elicits some recollection, the best current evidence suggests that it elicits less recollection than tasks that use intentional encoding and retrieval, such as matching.

4. VPC tests incidental recognition memory [e.g., 23,39,40]. Unlike matching, VPC does not require any explicit encoding, retention, or retrieval. Thus, whatever information is retrieved was incidentally encoded.

   This is possible. The concept of “incidental memory” is underspecified in the literature, but it is often described as similar to the concept of familiarity memory: an automatic process that encodes relatively vague information. There is robust evidence that the hippocampus is not necessary for familiarity memory [e.g., 15,33,but see 9], which is contrary to an ‘incidental memory’ account and the sensitivity of VPC to hippocampal damage. Empirically, incidentally encoded stimuli can support successful matching in both humans and monkeys [41-44]. Thus, an incidental encoding account raises many questions, but is not – in and of itself – a sufficient explanation of what VPC measures.

5. VPC performance measures both memory and other nonmnemonic factors [e.g., 45]. Thus, the lack of a viewing preference cannot be attributed cleanly to memory alone.

   We argue that this is almost certain. Monkeys show preferential viewing of different stimuli for different reasons, such as enjoyment vs information seeking, and these reasons can be behaviorally dissociated [46]. For example, preferential viewing based on differences in low-level perceptual features tends to be relatively stable across exposures, but preferential viewing based on differences in stimulus familiarity tends to dissipate with further exposure to the stimuli [46]. In addition to memory, it is likely that preferentially viewing is affected by multiple nonmnemonic factors, including novelty preference, habituation, motivation, and other processes. Next, we consider some of these nonmnemonic factors.

Influences on preferential viewing outside of explicit recognition

Classically, orienting to novel stimuli or contexts has been thought to reflect engagement of motivational systems that evolved to support survival [47]. As such, there are a number of physiological responses related to detection of novelty, including cardiac deceleration, altered skin conductance, and altered event-related potentials. These physiological changes are thought to enhance perceptual processing and to prepare for action. Although not typically considered as part of an ‘orienting response’, preferential viewing can be viewed this way. Here we discuss preferential viewing and its relationship with explicit measures of memory, returning to the themes of novelty, habituation, and motivation. To be clear, all of these are types of memory in their own right, in that they show the influence of past experience on current behavior. However, they all influence preferential viewing in ways that are separable from explicit tests of recognition, such as matching or yes/no reports.

Novelty.

Because the memory interpretation of VPC relies on novelty preference, VPC must also test the strength of an individual’s novelty preference. However, novelty preferences show large individual differences, low reliability, shifts during aging, and parameter-dependent effects, all in ways that do not track explicit measures of memory. Behaviorally, novelty preference to single items varies significantly among individual monkeys [48]. Indeed, although most VPC studies report a novelty preference [e.g., 20], some show a familiarity preference [e.g., 49,50]. In human infants, novelty preferences are typically low and only weakly correlated from one week to another [51]. In humans, subjects who explicitly remember the stimuli often do not show a novelty preference, instead often showing either no preference or a familiarity preference. For example, after having seen a simple action with puppets, infants showed no viewing preference for or against the puppets, but were able to show significant deferred imitation, demonstrating a dissociation of explicit memory and preferential viewing [52]. Other studies have found age-related differences; infants displaying a familiarity preference, moving to indifference as the children aged, and then finally to a novelty preference in older children [53-55]. In adults, some studies have found that viewing preference varies as a function of retention interval, with novelty preferences at short intervals, indifference at medium intervals, and familiarity preferences at long intervals [56]. Subjects showed accurate explicit recognition memory at all these intervals, regardless of viewing preference [56], demonstrating that novelty preference is independent of memory. These individual differences, reliability effects, age effects, and parameter-dependent effects have prompted caution when interpreting preferential viewing paradigms [e.g., 53,57]. Critically, failure to show a viewing preference should not always be interpreted as a lack of memory [e.g., 53,57].

Neurally, the network for detecting novelty and using novelty to guide exploration is distinct from that involved in recognition memory. In humans, fMRI dissociates novelty, recollection, and familiarity [58]. In monkeys, single unit recordings dissociate neurons that code stimulus novelty, familiarity, and recency [59]. Brain regions involved in detecting, responding to, and utilizing item novelty are still being mapped [60], but include the zona incerta [61], area TE [59,61-63], the entorhinal and perirhinal cortex [59,61], the amygdala [62,63], the basal ganglia [62,64,65], the hippocampus [66,67], and other areas. It is still unclear which of these areas are critical to different aspects of novelty processing and which receive novelty signals but are not critical. Notably, this novelty network partially overlaps with several common targets of investigation for memory studies.

Habituation.

It has been suggested that VPC tests some aspect of habituation [45]. Habituation is quantified as the speed with which subjects decrease responding to a stimulus. Habituation is dissociable from declarative memory and is usually classified as nondeclarative, rather than declarative like recognition [68]. For example, people look less at visual scenes upon repeated viewing – in this case, the scenes are not paired with another scene as in VPC – but this habituation of the viewing response is not related to explicit recognition of those scenes in control participants and is intact in patients with impaired declarative recognition due to medial temporal lobe damage [69]. Similarly, the habituated skin conductance response upon repeated exposure to images or sounds is equivalent in both control participants and patients who show impaired declarative memory due to medial temporal lobe damage [70]. Even within habituation, there are likely multiple processes that habituate at different rates, as evidenced by the different habituation rates of the various autonomic and physiological reactions to novel stimuli, such as heart rate, skin conductance, and event-related potentials [47]. The need to disentangle habituation and recognition has been raised by others in regards to studies of c-fos activation as a memory measure [71]. In short, VPC relies on habituation, but habituation is dissociable from explicit recognition.

Motivation.

The discrepancy between VPC and matching is partially due to motivation. Unlike VPC, matching makes a reward contingent on reporting memory. Reward magnitude can affect performance in a variety of tests such as short-term spatial memory [72], interval timing [73], or visual change detection [74]. In VPC, prior stimulus associations, such as seeing a stimulus just before seeing an unpleasant image, or the inherent aesthetic of some stimuli being more pleasant to view than others, can eliminate preferential viewing while leaving explicit recognition unaffected [75]. Motivation to look at a stimulus because of prior associations or inherent aesthetics competes with the motivation to look at it because it is familiar, whereas explicit choice is motivated primarily by external reinforcement – either a food reward or positive feedback – for following the matching rule. Thus, it is possible that a subject shows indifferent viewing due to the interplay of multiple conflicting motivations that control viewing but can still show accurate matching because of the unconflicted motivation to be reinforced.

In addition to novelty preference, habituation, and motivation, VPC is probably also affected by factors like subject hyperactivity, attentional demands, low-level perceptual features of the stimuli, and others. Table 1 presents a selected list of papers whose findings are inconsistent with VPC as a process-pure test of recognition memory.

Table 1.

Selected Findings Questioning the Validity of Novelty Preference as a Measure of Memory

Paper	Subjects	% Male	Task(s)	Notable result
[46]	Monkeys (Macaca mulatta)	100	VPC	Preferential viewing was independently affected by multiple motivations, such as pleasure or interest
[55]	Humans	50	VPC	Infants' preference for viewing familiar or novel items shifted as they aged
[54]	Humans	49	VPC	Infants' preference for viewing familiar or novel items shifted as they aged
[51]	Humans	Not reported	VPC	Individual infants' novelty preferences were largely unreliable from week to week
[23]	Monkeys (Macaca mulatta)	67	VPC vs DNMS	Hippocampal damage profoundly affected viewing preference yet left matching intact
[28]	Humans	35	VPC vs Yes/No	Preferential viewing was uncorrelated with matching accuracy
[3]	Monkeys (Macaca fascicularis)	100	VPC vs DNMS	Hippocampal damage had a larger effect on viewing preference than on matching
[52]	Humans	48	VPC vs deferred imitation	Infants demonstrated memory via deferred imitation despite indifferent viewing
[20]	Monkeys (Macaca mulatta)	47	VPC vs DNMS	Hippocampal damage profoundly affected viewing preference yet left matching mostly intact
[25]	Humans	0	VPC vs DMS	A patient with hippocampal damage showed robustly spared recognition across tasks, yet profoundly reduced viewing preference
[56]	Humans	23	VPC vs DMS	Humans exhibited high levels of accuracy on matching tests irrespective of delay but exhibited viewing preferences that were delay dependent and ranged from familiarity preferences to novelty preferences to indifferent viewing
[48]	Monkeys (Samiri sciureus)	25	Object novelty preference	Baseline novelty preferences vary considerably among individual monkeys
[75]	Humans	50	VPC vs DMS	Conditioning or inherent stimulus aesthetics affected viewing preference but not matching
[49]	Monkeys (Macaca fuscata)	100	VPC	Sometimes monkeys display a familiarity preference instead of a novelty preference
[76]	Rats (Long Evans)	0	NOP vs DNMS	Both systemic administration and intracranial infusion of estradiol into perirhinal and entorhinal cortex improve novelty preference but decrease matching performance
[77]	Rats (Long Evans)	0	NOP vs DNMS	Estradiol injection into perirhinal cortex improves novelty preference but decreases matching performance
[69]	Humans	40	Scene viewing vs Yes/No	Habituation of viewing to repeated scenes and memory for those scenes is uncorrelated
[27]	Humans	29	VPC vs Yes/No	Recognition explained only a third of variance in viewing preference
[78]	Rats (Long Evans)	100	NOP vs mDNMS	Rats showed accurate recognition on both tests, but scores on the two tests were uncorrelated
[30]	Rats (Long Evans)	100	NOP vs mDNMS	Rats with perirhinal cortex lesions showed indifferent novelty preference but accurate matching
[24]	Monkeys (Macaca mulatta, Macaca fascicularis, Papio anubis)	76	VPC vs DNMS	Meta-analysis shows that hippocampal lesions in monkeys affect viewing preference but not matching accuracy

Note: Papers are ordered chronologically by publication year. % Male refers to the percentage of subjects in the study that were male. VPC = visual paired comparison, DNMS = delayed nonmatching-to-sample, DMS = delayed matching-to-sample, Yes/No = yes/no recognition judgement, mDNMS = modified delayed nonmatching-to-sample, NOP = novel object preference

Asymmetrical interpretation

To be clear, VPC is affected by memory. If the stimuli are matched for content, visual complexity, emotional valence, balanced in their assignment as targets or lures, etc., then a group-level viewing preference must indicate memory of those stimuli. VPC’s sensitivity to hippocampal damage in primates even makes it a potentially-informative choice for researchers interested in assessing hippocampal integrity per se, or in exploring a contributory role of the hippocampus to recognition. However, because multiple factors affect performance, results from a VPC test must be interpreted asymmetrically. A subject might remember the item but show no viewing preference because they do not have a strong novelty preference. A given manipulation, such as a brain lesion, might leave memory intact but cause indifferent viewing because it blunts habituation speed. A group might show no overall viewing preference because it is composed of equal numbers of subjects that show novelty preferences and familiarity preferences. In other words, viewing preferences can be interpreted as memory, but indifferent viewing cannot be interpreted as a lack of memory. For the same reasons, differences in the degree of novelty preference cannot be interpreted as differences in memory strength. This asymmetrical interpretation makes VPC a better test of memory’s presence than its absence, and thus a poor choice for detecting memory impairment. For readers attempting to interpret an often-conflicting literature, these considerations suggest that evidence from explicit tests of recognition, such as matching, should be weighted more strongly than those from tests of preferential viewing.

Matching tests

Explicit matching tests have several benefits over preferential viewing paradigms, but also some limitations that must be kept in mind. Perhaps the strongest benefit is that matching tests make reinforcement directly contingent on the subject indicating their memory. In the context of animal experiments that rely on food reinforcements, failure to select the nonmatching item in a DNMS task can be confidently interpreted as a lack of memory because we are confident that the hungry animal would have followed the nonmatching rule, and retrieved the food, if it could have. Likewise, in the context of human experiments, it is reasonable to trust the participants to follow test instructions, such as a the matching or nonmatching rule, to obtain reinforcement (e.g., points, positive feedback). A second major benefit of matching tests is that researchers have worked out clever paradigms to use variants of these tests to dissociate the contributions of different memory processes, such as recollection vs familiarity [79-81], memory from an individual trial vs long term habit memory [82,83], and active working memory vs passive familiarity [84,85]. The ability to dissociate different processes within the same memory test holds great promise for future studies on the neurobiology of memory.

A major limitation of matching tests as they are used in animal studies is that they require extensive training. This makes conducting studies of matching harder, slower, and costlier. More importantly, it introduces the chance that subjects will inadvertently learn some accidental statistical regularity in how the correct food locations are pseudo-randomized or the low level perceptual features of the stimuli used as targets vs lures. These problems are minimized with rigorous experimental design, but it is impossible to anticipate every unknown, and examples exist of monkeys learning task rules other than those intended [e.g., 86,87]. A second worry is that the intentional encoding and retrieval of matching tests will encourage subjects to maintain information in active working memory rather than long-term memory. This is a valid concern under some task parameters; however, monkeys are less likely to default to a working memory strategy than are humans [85] and the use of working memory can be discouraged through long retention intervals, long lists of to-be-remembered items [5], or using cognitive distractors during the retention interval [4,84].

Concluding remarks

The two most common tests of recognition memory, preferential viewing and matching/nonmatching, provide different results. In this article, we made the case that preferential viewing paradigms do not provide a process-pure test of recognition memory. Instead, they likely test a combination of recognition, novelty preference, habituation, motivation, and other factors. Quantifying and disentangling these influences will require future studies that combine nuanced behavioral analyses with causal neural manipulations, are appropriately powered, and consider the need to evaluate factors like sex as a biological variable (see Table 1 and Outstanding Questions). Overall, failure to show preferential viewing cannot be cleanly interpreted as a lack of recognition memory. In considering the neurobiology of recognition memory, explicit matching tests should be preferred by memory researchers, and evidence from matching tests should be weighed more heavily than evidence from preferential viewing tests.

Outstanding Questions.

What percentage of the variance in preferential viewing is accounted for by different nonmnemonic factors, such as individual novelty preference, habituation, etc?

Some studies of human development report a shift from a natural familiarity preference, to indifference, and finally to novelty preference. Do nonhuman primates and rodents show a similar shift during development? If so, what is the time course?

Has the asymmetry in ease of conducting these tests created an asymmetric “file drawer” problem in which null results from one paradigm are less likely to be published?

Where in the brain are neural circuits shared and where are they distinct in implementing preferential viewing and matching tasks?

Recognition is generally characterized as consisting of a vague familiarity component and a detailed recollection component, but the relative contributions of these components have not been robustly compared between paradigms. Under matched conditions, do preferential viewing and matching/nonmatching elicit the same balance of familiarity and recollection?

In identifying the brain bases of recognition, the hippocampus is the most commonly studied structure and the one for which the most direct comparisons exist of preferential viewing and matching/nonmatching. How would animals with selective lesions of other structures important for memory, such as the perirhinal cortex, anterior thalamus, mammillary bodies, or retrosplenial cortex, perform on direct comparisons of preferential viewing and matching/nonmatching?

It is clear that preferential viewing measures a combination of mnemonic and nonmnemonic factors. It is also clear that hippocampal damage often spares recognition memory in tests that require explicit choice or in which human subjects can declare their memory. So, why, exactly, is preferential viewing uniquely sensitive to hippocampal damage?

Highlights.

The two most common paradigms for testing recognition memory in nonhuman primates provide conflicting evidence about its underlying neurobiology.

The most widely used paradigm, preferential viewing, likely confounds memory with nonmnemonic processes like novelty processing, habituation, motivation, and others.

Despite widespread use, preferential viewing correlates poorly with more traditional measures of explicit recognition.

We propose that memory paradigms that require explicit choices should be preferred by researchers and weighted more heavily by readers.