It is controversial whether monkeys, like human subjects, can recall, upon instruction, specific information about an event in memory. We therefore tested macaque monkeys on a task that was originally developed to study such active controlled memory retrieval in human subjects and we were able to demonstrate that monkeys, like human subjects, can retrieve, upon command, specific components of previously encoded events. Furthermore, following earlier functional neuroimaging work with human subjects showing the mid-ventrolateral prefrontal cortex to be involved in such active controlled retrieval, we recorded single-neuron activity within this region of the monkey brain while the monkeys performed the active retrieval task. Neuronal responses were related to the retrieval and the decision whether the retrieved information was the instructed one. These findings demonstrate, for the first time, an impressive capacity by macaque monkeys for controlled memory retrieval and, in addition, provide neurophysiological evidence about the role of the mid-ventrolateral prefrontal cortex in such controlled retrieval.
Macaque monkeys demonstrate an impressive range of mnemonic abilities. In the classic delayed non–matching-to-sample test (and the related matching-to-sample test), monkeys demonstrate that they can recognize whether a test stimulus has been previously experienced or not, that is, whether it is familiar or novel (Mishkin and Delacour 1975). Monkeys can also learn, as a result of many training trials, strong, fixed, and unambiguous relations between pairs of stimuli, such as stimulus–stimulus and stimulus–motor associations acquired in conditional associative learning tasks (Halsband and Passingham 1982; Petrides 1982; Petrides 1985; Sakai and Miyashita 1991; Naya et al. 1996), the various parts that comprise a complex stimulus in compound-stimulus tasks (Murray et al. 1993), or between stimuli and specific contexts in scene-learning tasks (Gaffan 1994). In the working memory domain, monkeys can code and recall constantly changing locations of stimuli, as in the classic delayed-response and delayed-alternation tasks (Jacobsen 1936; Fuster 1973; Goldman-Rakic 1995; Goldman-Rakic and Leung 2002). They can also track in working memory their recent choices from a set of stimuli (Petrides 1991a, 1995), track whether particular objects from an expected set have or have not recently been presented (Petrides 1995), and recall the precise serial order of occurrence of a small set of visual stimuli after having seen them only once (Petrides 1991b). These working memory tasks have been used successfully to investigate various aspects of the role of the prefrontal cortex in working memory (e.g., Fuster 1973; Funahashi et al. 1989; Petrides 1991a, b; Funahashi et al. 1993; Goldman-Rakic 1995; Goldman-Rakic and Leung 2002; Petrides 2005). Indeed, many of these tasks were developed specifically to address questions regarding the role of the prefrontal cortex in working memory.
There is, however, no evidence whether nonhuman primates can engage successfully in the active controlled retrieval of information from memory, that is, whether they can retrieve, upon instruction, specific aspects of the information embedded in a multicomponent event in memory, especially in the context of high interference from many other similar events. Certainly, monkeys can recognize that they had seen an object before (e.g., in the delayed nonmatching-to-sample tasks) and that an object was associated with certain other objects or places. However, as Griffiths et al. (1999) argued, adequate performance on the delayed nonmatching-to-sample task can be explained on the basis of familiarity discrimination and, in fixed-pair tasks, by the learning of fixed stimulus–stimulus or stimulus–response associations (e.g., in compound-discrimination tasks) or associating a unique what–where configuration with reward on the basis of many experiences (e.g., in scene-learning tasks). Can monkeys isolate from their memory and report a specific component of an event? This question is critical because there is considerable evidence from functional neuroimaging studies that a certain part of the prefrontal cortex, the mid-ventrolateral region, plays a major role in active controlled retrieval from memory (e.g., Cadoret et al. 2001; Kostopoulos and Petrides 2003; Amunts et al. 2004; Badre et al. 2005). Although functional neuroimaging studies in normal human subjects can indicate the engagement of a cortical area in a particular cognitive process by demonstrating changes in blood flow within it in relation to that cognitive process, they cannot provide direct information on the neuronal activity that underlies the cognitive process. Such detailed exploration of the neural computations in the ventrolateral prefrontal cortex can only be carried out on nonhuman primates. The aim of the present investigation, therefore, was 2-fold: to develop a paradigm to assess active controlled retrieval in macaque monkeys and then to use this paradigm to carry out an initial exploration of neuronal activity related to active retrieval by recording the activity of single neurons in the mid-ventrolateral prefrontal cortex.
Information in memory is embedded in a network of associative relations between stimuli or between stimuli and specific contexts, which can, in turn, be thought of as complex stimuli. At one extreme, the relations between various stimuli (or stimuli and contexts) may be strong, stable, and unambiguous (e.g., stimulus X is always or often linked with stimulus Y) as in the compound-discrimination and scene-learning tasks mentioned above. In such cases, the evocation of one stimulus triggers automatically the retrieval of the associated one. For instance, if car A is always or often parked in location X and car B is always or often parked in location Y, retrieval of the location X when I think of car A is quasi-automatic: the image/memory of car A triggers through strong associations location X. At the other extreme, the stimuli may be embedded in unstable and ambiguous relations that vary across trials requiring a considerable degree of active controlled retrieval (Petrides 2002, 2005). For instance, across days, cars A and B may be parked in locations X and Y randomly and with equal probability and, therefore, the location of the cars cannot be predicted on the basis of well-learned associations that are stable across days. Retrieval from memory in such a case requires the disambiguation of the relations between stimuli across trials and the capacity to isolate the required feature from one specific event. For the purpose of this study, we adapted for the monkey a task that we had originally developed for human subjects to provide an operational definition of active controlled retrieval in order to study cerebral activation related to such retrieval with functional neuroimaging (Cadoret et al. 2001; Kostopoulos and Petrides 2003). Briefly, this task was as follows. Various stimuli (e.g., 4 objects and 4 locations or 4 shapes and 4 colors) would occur in different combinations during the encoding phase of a series of trials. Importantly, across trials, any 1 of the 4 objects would appear in any 1 of the 4 locations (or any 1 of the 4 shapes with any 1 of the 4 colors), randomly but with equal probability, ensuring that no stable relations could be established between particular objects and particular locations. The principle of the retrieval task was to present, during the encoding phase of each trial, one combination of the stimuli (e.g., object A in location X) and, after a short delay, to present a cue instructing the monkey to retrieve one or the other aspect of that particular encoding event (e.g., the particular object or the location in the preceding encoding event) (Fig. 1A). Finally, there was a test phase during which one of the objects (A, B, C, or D) was presented in one of the locations (X, Y, Z, or W) and the monkey had to decide whether a part of this test stimulus (e.g., object B in location X) corresponded to the particular aspect of the encoded event that the monkey was instructed to retrieve (e.g., object A or location X in the example above) (Fig. 1A). Thus, the correct response on the test phase of the trial was entirely dependent on the monkey's ability to retrieve the specific feature of the stimulus in memory that the retrieval cue had instructed.
Materials and Methods
Two adult male monkeys (Macaca fascicularis), weighing 7 and 8 kg, were used for this experiment. Training and recording procedures were approved by the Animal Care Committee of the Montreal Neurological Institute, McGill University.
Each trial was initiated by the appearance of a fixation point at the center of the computer screen (Fig. 1A). The monkeys then had to press a response key and keep pressing it until the test phase of the trial; 750 ms after the monkey had pressed the response key, the encoding phase began and, for Monkey 1, 1 of 4 objects was presented for 750 ms in 1 of 4 locations around the fixation point. Thus, the encoded stimulus had 2 characteristics: object and location. For Monkey 2, during the encoding phase, 1 of 4 shapes in 1 of 4 colors appeared at the center of the screen. Thus, for Monkey 2, the 2 attributes of the encoded stimulus were shape and color. The encoding phase was followed by a short delay (500 ms) during which the screen was blank, and then a cue appeared (for 900 ms) instructing the monkey to retrieve a specific aspect of the information presented during the encoding phase. For Monkey 1, a green cue instructed retrieval of the object that was presented during the encoding phase, whereas a blue cue instructed retrieval of the location. For Monkey 2, a gray cue instructed the animal to retrieve the shape of the stimulus, whereas a cue made of patched colors instructed retrieval of the color of the stimulus. After a further delay (1000 ms) in which the screen was blank, the test phase followed. In the test phase, a stimulus was presented that again consisted of 2 parts: an object and a location for Monkey 1 and a shape and a color for Monkey 2. The monkey had to decide whether a part of this test stimulus (its object feature or location, shape or color) corresponded to the attribute (e.g., the location) of the previously encoded stimulus that the cue had instructed the monkey to retrieve. If the relevant part of the test stimulus (e.g., its location) matched the particular aspect of the encoded stimulus that the cue had instructed the monkey to retrieve (e.g., its location), the monkey had to release the response key as soon as possible (within 900 ms) thus essentially reporting: “yes, the relevant part of the test stimulus matches the attribute of the encoded stimulus retrieved from memory.” The release of the response key resulted in the screen becoming black, indicating termination of that trial. If the relevant part of the test stimulus did not match the particular aspect of the encoded stimulus instructed by the retrieval cue, the monkey had to keep pressing the response key until the screen became black (1200 ms later). By continuing to press the response key, the monkey was reporting: “no, the relevant part of the test stimulus does not match the attribute of the encoded stimulus retrieved from memory.”
For both monkeys, the retrieval trials described above were randomly intermixed with no-retrieval trials. These trials were identical to the retrieval trials in terms of the initial presentation of the stimulus information (i.e., the encoding phase), but the cue that followed did not instruct the retrieval of any attribute of the encoded stimulus, but simply instructed the monkey to wait for the impending test phase. In the test phase of the no-retrieval trials, a stimulus composed of 2 parts appeared and the monkey had to release the response key if both parts were identical, but to keep pressing the response key until the screen became black if the 2 parts of the test stimulus were different. An important point to note is that the encoding phases of the retrieval and no-retrieval trials were identical, and thus, the monkeys did not know whether they would be required to retrieve or not until the occurrence of the instruction cue; consequently, the encoding demands of all trials were identical. The retrieval and no-retrieval trials occurred in a pseudorandom order. A typical series (random series) was composed of 24 trials, with 16 retrieval and 8 no-retrieval trials. For Monkey 1, in 8 of the retrieval trials, object retrieval was instructed and in the other 8 trials location retrieval was instructed. For Monkey 2, in 8 of the retrieval trials, shape retrieval was requested and in the other 8 retrieval trials color retrieval was instructed. Correct trials were rewarded with apple juice.
A recording chamber (18-mm internal diameter) was stereotaxically implanted over the exposed ventral prefrontal cortex under sodium pentobarbital (30 mg/kg) anesthesia. Recordings were conducted in both hemispheres in Monkey 1 and in the left hemisphere in Monkey 2.
After the monkeys had recovered from surgery, recording sessions were conducted on a daily basis. Glass-insulated tungsten microelectrodes (impedance 0.5–1.7 MΩ, measured at 1 kHz frequency) were inserted through the dura into the brain with a hydraulic microdrive attached to an X–Y micropositioner. Cell search was conducted while the animal performed the retrieval task. Neuronal activity was monitored on an oscilloscope, and individual action potentials were isolated with a voltage discriminator. Each isolated neuron was tested with the task on one or multiple series of trials.
Before the end of the recordings, electrolytic lesions were made by passing direct current through the recording microelectrode (20 μA, 20 s) at selected locations. These lesions were used to reconstruct the penetration sites and to establish architectonic areas within which the recording sites were located. A few days later, the monkeys were deeply anesthetized with an overdose of sodium pentobarbital (100 mg/kg intravenous) and perfused transcardially with a solution of 4% paraformaldehyde in cacodylate buffer. The frontal lobe of the brain was blocked, frozen, and sectioned at 40 μm. The coronal sections were stained with cresyl violet for architectonic analysis.
Analysis of neuronal discharges was performed on correct trials. The neuronal discharge for each cell was analyzed in 6 task periods: 1) the “baseline period”: the last 500 ms of the fixation period before stimulus presentation in the encoding phase; 2) the “encoding period”: 750 ms during which the visual stimulus complex was presented; 3) the “first delay period (Delay 1)”: 500 ms after the stimulus was presented; 4) the “instruction cue period”: 900 ms; 5) the “second delay period (Delay 2)”: 1000 ms after the cue presentation and before the test phase; 6) the “test period.” The number of spikes during each period was normalized by the duration of each period (i.e., the number of spikes in each period was divided by the duration) and calculated as the spike rate (spikes per second). A neuron was judged as having task-related activity for a specific period if the spike rate during this period, at least on 2 consecutive bins (20 ms), deviated by 3 standard deviations (SDs) from the mean spike rate that was calculated during the baseline period.
In order to examine whether the activity of each neuron was modulated by the retrieval events, we analyzed the firing rate during the instruction cue period, the second delay after the cue, and the test period. Because of the length of these periods, we divided them into early and late phases. The early and late instruction cue periods were 450 ms each, and the early and late second delay periods were 500 ms each. The test period in the case of the “no” (i.e., nonmatching) responses that required maintenance of the key pressing for 1200 ms was divided into the early phase (first 600 ms) and the late phase (second 600 ms). In the case of the “yes” (i.e., matching) responses that required the release of the key, the test period ended with the release (400 ms ± 49). We performed for each recorded neuron a 2-way analysis of variance (ANOVA; P ≤ 0.05) in which the factors were: periods (as defined above) and conditions (no-retrieval, object retrieval, and location retrieval for Monkey 1 or no-retrieval, color retrieval, and shape retrieval for Monkey 2). If the ANOVA indicated a significant interaction, the Newman–Keuls test was used to explore further the significant interaction.
In the analysis of the results for the test phase, we always separated, initially, the trials in which the correct decision was indicated by releasing the response key or maintaining the key press, that is, the type of motor response was entered as a separate factor in the ANOVA. A neuron was considered to exhibit retrieval-related activity during the test phase only if the neuronal firing could not be explained fully by the motor response. In other words, we did not consider a neuron to exhibit retrieval-related activity in the test phase if the main effect of response type (i.e., release or maintain the key press) was significant and there was no significant interaction with the retrieval conditions.
Behavioral Results: Evidence for Active Controlled Memory Retrieval
Note that in this task, the encoded stimulus is a combination of 2 characteristics: a specific object in a particular location or a specific shape in a particular color. Note also that, after the encoding event, the monkey is instructed to retrieve 1 of the 2 characteristic features of the encoded stimulus complex. In the test phase of the retrieval task, in a random half of the trials, the test stimulus was a compound that was either identical in both characteristics (object and location or shape and color) to the encoded stimulus or different in both aspects. For instance, if object A was presented in location X during the encoding phase, object A in location X could be presented in the test phase (identical in both characteristics) or object B in location Y might be presented (different in both characteristics). We call these trials type A trials. In the remaining random half of the test trials, which we call type B trials, only one of the dimensions of the test stimulus was the same as that of the encoded stimulus. In the example above, object A might appear in location Y or object B might appear in location X. Note that the 2 examples of trial shown in Figure 1A are type B.
The animals were trained on this task over a period of 18 months. During this period, the animals were first trained to press or release the response key and to attend to stimuli on the screen before training on the specific task. To examine the retrieval performance of the 2 monkeys, the last 10 sessions of the training period were selected. At that time, the performance of the 2 monkeys had reached a plateau. The mean rates of success have been calculated when monkeys were clearly performing above chance (50%). Figure 1B shows the mean rate of success for type A, type B, and no-retrieval trials for Monkey 1, who performed the object/location retrieval trials, and for Monkey 2, who performed the shape/color retrieval trials. ANOVA showed significant differences in performance across the different types of trial both for Monkey 1 (F = 7.25, degree of freedom [df] = 2, P < 0.001) and for Monkey 2 (F = 84.31, df = 2, P < 0.001). Further analysis of this effect with the Newman–Keuls test showed that, for Monkey 1, performance on type A and type B trials was significantly inferior to performance on the no-retrieval trials and for Monkey 2, performance on type B trials was significantly inferior to performance on type A and no-retrieval trials. Thus, as would be expected, performance on the test phase of the no-retrieval trials in which the decision was made on the basis of the perceptual features of the test stimulus (see Behavioral Methods) was significantly better than performance on the retrieval trials in which the decision was based on retrieved information from a memory representation. Note also that type A trials were, on average, easier than type B trials because both characteristics of the test stimulus (object and location, shape and color) were the same or different from those of the sample stimulus. By contrast, on the more challenging type B trials, only one aspect of the test stimulus (object or location, shape or color) matched the encoded stimulus and the monkey had to decide whether the test stimulus matched the encoded stimulus in the characteristic that was instructed by the retrieval cue. For example, if the monkey saw a “blue house” during the encoding phase of a particular trial (as in Fig. 1A, Monkey 2) and was instructed to retrieve the color of the encoded information, the monkey should release the response key if the test stimulus was a “blue star” (thus stating “yes, the color of the test stimulus matches that of the encoded stimulus”), but should keep pressing the key if the test stimulus was a “green house” (thus stating “no, the color of the test stimulus is not the same as the color of the encoded stimulus”). By contrast, if the instruction was to retrieve the shape of the encoded stimulus (i.e., the “blue house” in the example above), the monkey should release the key if the test stimulus was a “green house” (thus stating “yes, the shape of the test stimulus matches that of the encoded stimulus”), but should keep pressing the key if the test stimulus was a “blue star” (thus stating “no, the shape of the test stimulus is not the same as the shape of the encoded stimulus”). Performance above chance on type B trials indicates that the monkeys are able to isolate from the stored information (e.g., blue house) the relevant dimension (e.g., blue) that the retrieval cue had instructed in order to make the correct decision in the test phase of the trial. Thus, the correct response during the test phase was entirely dependent on the correct retrieval of the instructed stimulus dimension.
During the test phase of the trials, if the test stimulus matched the particular aspect (object or location, shape or color) of the encoded information that had been instructed by the retrieval cue, the monkey had to release the response key (thus stating “yes, there was a match along the instructed dimension”) as soon as possible. There were significant differences in reaction times for the various types of trial (Fig. 1C) (F = 38.48, df = 2, P < 0.001 for Monkey 1; F = 15.63, df = 2, P < 0.001 for Monkey 2). Further analysis with the Newman–Keuls test showed that the reaction times on correct type B trials were slower than on correct type A trials, for both monkeys. This finding was expected because in type A trials both features (object and location, shape and color) of the test stimulus matched or did not match those of the encoded stimulus complex, whereas in type B trials the required feature (object or location, shape or color) isolated from the memory representation matches only one aspect of the encoded stimulus complex.
On a large proportion of type B trials (76.5% for Monkey 1; 77.5% for Monkey 2), the reaction times were within 2 SDs of the mean reaction times on type A trials. This capacity to respond on a large proportion of type B trials within the range of type A trials suggests that monkeys had already retrieved the specific aspect of the information (e.g., color) instructed by the retrieval cue before the test stimulus was presented and were, therefore, not perturbed, at the test time, by the irrelevant dimension. On the remaining proportion of type B trials (23.5% for Monkey 1; 22.5% for Monkey 2) on which the monkeys were clearly slower than on the type A trials, the monkeys had probably not started retrieving the required information until the test stimuli were presented or they retrieved the required information again during the test phase.
Neuronal Discharges Related to Active Retrieval
A sample of 117 cells was recorded in the mid-ventrolateral prefrontal cortex (Fig. 2) during the performance of the retrieval task and 62% of these neurons were task related. Of the task-related neurons, 51% demonstrated a significant change in their firing rate in relation to one aspect of active memory retrieval, that is, a significant difference in the neuronal firing rate between the retrieval and the no-retrieval trials or between the different types of retrieval trials (object or location, shape or color) during one or more of the retrieval periods: the instruction cue, the postinstruction delay 2, and the test periods. In both monkeys, retrieval-related neurons were recorded in the mid-ventrolateral prefrontal cortex, namely areas 45 and 47/12 (Fig. 2A). It has been shown that these 2 areas of the macaque monkey ventrolateral prefrontal cortex correspond, architectonically, to comparable areas in the human ventrolateral prefrontal cortex (Petrides and Pandya 2002; Petrides et al. 2005), where increased activity related to active retrieval has been reliably observed in functional neuroimaging studies (Cadoret et al. 2001; Kostopoulos and Petrides 2003). Area 45 in the human and macaque monkey brain is characterized by a well-developed granular layer IV and clusters of very large neurons in the deeper part of layer III and moderately large neurons in layer V (see Fig. 2B,C). Thus, there is a clear asymmetry in the size of the largest neurons in layer III compared with those in layer V. Area 47/12, which lies ventral and anterior to area 45, lacks the very large neurons in layer III that are conspicuous in area 45 (see Fig. 2C). We should emphasize here that we define area 45 of the monkey by the architectonic criteria of area 45 in the human ventrolateral prefrontal cortex. This is important because, in several macaque monkey studies in the past, a part of the ventral prearcuate cortex close to the sulcus principalis was termed “area 45” and was linked to oculomotor function. This part of the prefrontal oculomotor cortex belongs architectonically to ventral area 8 and is not included in area 45 as defined in the present study which uses the architectonic characteristics of area 45 of the human ventrolateral prefrontal cortex to define area 45 in the monkey (for further discussion of this issue see Petrides and Pandya 2002; Petrides et al. 2005). The part of the oculomotor cortex that, in the past, had been termed as area 45 (and which we and other investigators consider to be part of area 8) has very large neurons in layer V (Stanton et al. 1989), which is clearly not the case for macaque area 45 when defined by the same criteria as those used to define area 45 in the human prefrontal cortex (see Fig. 2C). Area 45 extends from the inferiormost part of the lower branch of the arcuate sulcus (below area 8Av) for a considerable distance anteriorly. Its rostralmost border is often marked by a small dimple, the infraprincipal dimple, just below the sulcus principalis (Fig. 2C). Figure 2B shows an electrode tract from the present study that is located in area 45. Interestingly, we noted that the probability of finding retrieval-related neurons was higher in the most ventral part of area 45 than in its dorsal part (see Fig. 2A).
Retrieval-Related Neurons: Cue Phase
In the present memory task, the retrieval process was initiated by an instructional cue that indicated retrieval of a specific feature of the previously encoded stimulus complex. Note also that there were trials during which the cue indicated no retrieval. Of the task-related neurons, 49% exhibited cue-related activity. About two-thirds of the neurons that modified their firing rate during the cue period (67%) had discharge rates that were not significantly different between the different conditions (nondifferential cue-related neurons). These neurons appeared to signal simply the initiation of that particular epoch of the trial. Of the cue-related neurons, 32% responded differentially according to the type of retrieval that was to be performed by the monkey (differential cue-related neurons). Half of these differential cue-related neurons changed their discharge in relation to cues indicating retrieval of specific aspects of the encoded stimulus complex (e.g., shape), whereas the other half were related to the cue that indicated that there would be no retrieval on that specific trial. These differential cue-related neurons appeared, therefore, to signal the initiation or noninitiation of the retrieval of specific information. Figure 3A shows an example of a neuron that responded only to the cue signaling no retrieval. Figure 3B shows a neuron that responded to the cues signaling retrieval and, more strongly, to the cue signaling retrieval for a particular feature (i.e., the shape) of the encoded stimulus.
Retrieval-Related Neurons: Delay 2 Phase
Some of the task-related neurons (15%) changed their firing rate during the delay that followed the cue period and preceded the test period, that is, delay 2 (see Fig. 1A). Figure 4A shows a neuron that increased its firing rate 450 ms before the appearance of the test stimulus in the retrieval trials compared with the no-retrieval trials (differential delay-related neuron). For the neuron shown in Figure 4B, the increase in activity prior to the appearance of the test stimulus was specific to the type of retrieval that had to be performed: there was a significant increase in activity during the latter part of the delay when location retrieval had been instructed than when no retrieval had been instructed. The instruction to retrieve the object did not result in greater activity during the latter part of the delay compared with the no-retrieval condition.
It is interesting to note that this increase in activity during the delay period was never observed in the no-retrieval trials, suggesting that this change in firing rate before the test phase is not signaling general anticipation of the test phase, but rather it is specifically related to the retrieval process. In addition, it should be emphasized that there was no increase in the activity of these neurons during delay 1 that occurred after the encoding phase but before the cue signaling retrieval. Thus, the increase in activity observed in delay 2 (which followed the instruction cue) was specifically related to the retrieval process. It could represent retrieval of the relevant information for the upcoming matching/nonmatching decision to be made in the test phase.
Retrieval-Related Neurons: Test Phase
During the test phase of the present task (see Fig. 1A), the monkey released the response key to indicate that the test stimulus on the screen matched the particular feature (e.g., the particular object) of the previously encoded information that had been instructed by the retrieval cue; in contrast, the monkey continued pressing the response key to indicate that the test stimulus did not match the particular feature of the memorized stimulus instructed by the cue. Thus, the matching/nonmatching decision was correlated with the motor response. The type of motor response was, therefore, always entered as a separate factor in the ANOVA used to analyze differential activity and a neuron was considered to exhibit retrieval-related activity during the test phase only if the neuronal firing was not accounted by the motor response (i.e., release the key or keep pressing the key). A high proportion of the task-related neurons (78%) modulated their discharge rate during the test period. We report here a striking type of neuronal response, observed during the test period, which correlated with the cognitive retrieval decision by the monkey, namely that the test stimulus contained the specific feature of the memorized information (e.g., the object or location) that was instructed by the cue.
Cognitive Disambiguation Neurons (Active Controlled Retrieval Decision Neurons)
Figure 5 shows a type of neuronal response observed during the test phase of the present retrieval task. Note that, during the test phase, this neuron was more active in trials in which location or object retrieval had been instructed than in trials in which no retrieval had been instructed (Fig. 5A). This neuron, therefore, was significantly more active during the test period but “only when the monkey was correctly deciding” that the test stimulus matched or did not match the instructed aspect of the previously encoded information (Fig. 5A). The neuron did not increase its firing rate in the test phase of the no-retrieval trials when the monkey was correctly making the same matching/nonmatching decision but now based on the perceptual features of the presented test stimulus and not on a comparison of the test stimulus with information in memory. Thus, the activity of this neuron during the test phase was coding the correct decision, which was based on the retrieval of a specific feature of the information in memory that was instructed by the retrieval cue. Note that the activity presented in Figure 5A during the test phase is mixed for matching and nonmatching decisions because there was no significant difference in activity between these decisions (see Fig. 5B). Thus, the activity of the neuron during the test phase was not related to the matching/nonmatching decision and, by implication, it was not related to the motor response that was correlated with the decision (see Materials and Methods).
It is also important to emphasize that the increase in neuronal response during the test phase when object or location retrieval had been instructed (Fig. 5A) was not related to the reward because the analysis is based only on correct trials and all trials (i.e., retrieval and no-retrieval trials) ended with the exact same reward. Furthermore, the retrieval-related increase in the firing of neurons such as the one illustrated in Figure 5 cannot be attributed to the perception of the stimuli for a number of reasons. First, as can be observed in Figure 5C, the neuron increased its firing rate during the test phase when the retrieval decision was made (i.e., the decision whether the test stimulus was or was not the one that had to be retrieved from memory), but the neuron did not respond during the encoding phase when the exact same stimuli were shown in order to be encoded. Thus, the differential response during the test period when retrieval was required in comparison with the encoding phase during which the exact same stimuli were presented further supports the interpretation that the firing of these neurons is related to the memory retrieval decision (Fig. 5C). Second, the retrieval-related firing during the test phase was not stimulus specific (i.e., it was not related to a particular stimulus), but rather to the cognitive retrieval decision that a feature of the particular test stimulus (e.g., its location) corresponded to the specific feature of the memorized stimulus that had to be retrieved (e.g., location). Thus, this neuronal response in the test phase cannot be related to the perception of the presented test stimuli, but rather to the cognitive decision made on the basis of the retrieved memory information.
Given that perceptual, reward, motor, and matching/nonmatching decision factors can be ruled out, the increased activity of this neuron during the test period when retrieval was required (compared with the test period when no retrieval was required) must be related to the active retrieval decision. Note that the correct response on the test phase of the retrieval trials in the present task was “entirely dependent on” the monkey's ability to decide, “based on the instruction cue,” that a part of the test stimulus (e.g., its location) is or is not the same as the corresponding part (e.g., location) isolated from the encoded compound stimulus stored in memory. As can be seen schematically in Figure 5D, in the test phase, exactly the same stimulus complex may be in memory on 2 different trials (e.g., object A in location 3 on trial n and trial n + 1) and the same test stimulus on the screen (e.g., object A in location 4 on trial n and trial n + 1), but the monkey must make a completely different decision depending on the instruction cue, that is, whether to retrieve the object (O) or the location (L). If the instruction cue is to retrieve the object (e.g., trial n in Fig. 5D), the correct decision is that there is a match in the relevant feature between the test stimulus on the screen and the stimulus in memory. By contrast, if the instruction cue is to retrieve the location (e.g., trial n + 1 in Fig. 5D), the correct decision is that there is no match between the test stimulus on the screen and the stimulus in memory. Thus, the increased response during the test phase in the retrieval trials cannot be related to a general matching/nonmatching decision per se but rather to the cognitive decision that a part of the test stimulus corresponds or not to the specific feature isolated from memory. Note that the firing of this neuron reflecting the correct decision must be based on the integration of information that is in memory (i.e., the encoded stimulus), on the screen (i.e., the test stimulus), and the instruction cue (see Fig. 5D).
The response of some controlled retrieval decision neurons was greater when the monkey decided that there was a match between the relevant feature of the test stimulus and the stimulus in memory, whereas for other controlled retrieval decision neurons the response was higher when the monkey decided that there was no match between the relevant dimension of the test stimulus and the encoded information. In other words, there was a significant interaction between the type of decision (matching/nonmatching) and the type of retrieval (object/location). Figure 6 shows the activity of such a neuron, which had a higher response rate during a matching decision but was more responsive when the instruction was to retrieve the object feature in comparison with retrieval of the location. Thus, the activity of this disambiguation neuron is modulated by the type of feature that had to be retrieved (i.e., object or location).
We believe that this neuronal firing, which reflects the cognitive decision that a part of the test stimulus matches a specific feature isolated from a memory representation upon instruction, is a fundamental neurophysiological correlate of active controlled memory retrieval. It enables disambiguation between 2 or more situations in which, with the same information in memory, the decision about a particular test stimulus will be different depending on what the subject was instructed to retrieve. We therefore refer to this type of neuron as the “controlled retrieval decision neuron” or the “cognitive disambiguation neuron.” The response of these cognitive disambiguation neurons must not be confused with the matching or nonmatching neurons that had previously been reported in the literature (e.g., Fuster 1973, 2000; Goldman-Rakic 1995). Those neurons responded differentially based on whether the current stimulus on the screen matched (or did not match) the stimulus in memory and therefore could not disambiguate between 2 test situations in which the same stimulus is in memory (e.g., object B in location 3 on trial n and trial n + 1) and on the screen (e.g., object B in location 4 on trial n and trial n + 1), but the decision is different depending on what type of retrieval the instruction cue had initiated. The cognitive disambiguation neurons discovered in the present study can disambiguate (differentiate) between these 2 situations because the firing of these neurons is the result of the integration of a) the stimulus in memory, b) the test stimulus, and c) the retrieval instruction (see Fig. 6).
Neuronal Response in Easy (Type A) versus Difficult (Type B) Active Retrieval Decisions
Note that, in the task studied here, the encoded stimulus had 2 components (e.g., an object and a location) and that, after the instruction to retrieve one aspect of the encoded stimulus (e.g., the object), the monkey was faced with a test stimulus that also consisted of 2 components (e.g., object and location). The monkey now had to decide whether the relevant component of the test stimulus (e.g., object D in the test phase illustrated in the upper part of Fig. 7) matched the corresponding component of the encoded stimulus (e.g., object C) that had been retrieved based on the instruction (O). On some test periods, the test stimulus matched (or did not match) the encoded stimulus in terms of both components (Type A), whereas, on other test periods, only one component of the test stimulus matched the encoded stimulus (Type B). The decision was easier on Type A test periods compared with Type B test periods because the monkey did not have to overcome the influence of the irrelevant dimension (see Behavioral Results: Evidence for Active Controlled Memory Retrieval for Type A vs. Type B trials). This difference in the difficulty of the cognitive decision was reflected in the firing rate of the controlled retrieval decision neurons, that is, the neuron responded significantly more to Type B than to Type A test phases. Figure 7 shows the response of one such neuron when retrieval of the object dimension of the encoded stimulus had been instructed. Note that this neuron was a nonmatch retrieval decision neuron, that is, it responded when a nonmatch decision had to be made. In the example of the test phase displayed in the upper part of Figure 7, the test stimulus (e.g., object D in location 4) differs from the stimulus in memory (e.g., object C in location 3) both in terms of the feature that the monkey had been instructed to retrieve (e.g., object) and the noninstructed feature (e.g., location; Fig. 7, Type A), whereas, on another trial, the test stimulus differs from the stimulus in memory only in terms of one feature (e.g., object but not location; Fig. 7, Type B). Although the correct decision in both cases is that the test stimulus does not match the component of the stimulus in memory that the monkey had been instructed to retrieve (i.e., object C), the decision in the latter case (Fig. 7, Type B) is harder because the monkey has to overcome the fact that the test stimulus matches the encoded stimulus in terms of location (i.e., the noninstructed dimension). Note that, although the neuron responds in both test phases reflecting the correct nonmatch decision, the neuronal response is greater in the more difficult decision (i.e., Type B). Thus, the firing rate of the controlled retrieval decision neurons is modulated by the difficulty of the memory decision, providing further support to the idea that these neurons are engaged in computations necessary to make the correct cognitive judgment.
The findings of the present study demonstrate, for the first time, an impressive capacity by macaque monkeys for active controlled memory retrieval and, in addition, provide initial neurophysiological evidence about the role of the mid-ventrolateral prefrontal cortex in such retrieval. In this task, 4 objects and 4 locations (or four shapes and four colors) were paired with each other randomly and with equal probability across trials, and therefore, there were no fixed, stable, and unambiguous associations between the stimuli acquired over many successive trials (e.g., a particular location always being associated with a particular object) that could automatically trigger strongly associated memory representations (e.g., the memory of location X in response to the presentation of object A). The pairings between the 2 components that comprised the encoding event (object/location, shape/color) varied from trial-to-trial, that is, the same stimuli were paired randomly and equiprobably with all the other stimuli across trials, and therefore, there was considerable interference between the similar memory representations established by the encoding events across trials. The monkey had to retrieve, upon instruction, a specific component from an encoding event. Thus, retrieval from memory required the isolation, upon instruction, of a specific feature (e.g., the particular object or particular location) from a recently encoded event. The results showed that monkeys, like human beings, are capable of isolating, upon instruction, and reporting specific aspects of a representation stored in short-term memory. Thus, they exhibit a level of memory performance that goes far beyond the learning of fixed stimulus–response or stimulus–stimulus associations that is demonstrated in conditional associative (Halsband and Passingham 1982; Petrides 1982, 1985; Sakai and Miyashita 1991; Naya et al. 1996), compound-stimulus (Murray et al. 1993), and scene-learning (Gaffan 1994) tasks, as well as the recognition of previously encountered stimuli based on familiarity that is examined in the delayed nonmatching-to-sample task (Mishkin and Delacour 1975).
Earlier work had demonstrated that one species of birds, the scrub jays, could display, implicitly, knowledge of an event (i.e., episodic-like memory) in their naturalistic food-caching behavior (Clayton and Dickinson 1998). However, there is no evidence that this finding is anything more than a demonstration of the existence of implicit knowledge of an event tightly linked to a specific naturalistic behavioral repertoire (Griffiths et al. 1999). Thus, the presently demonstrated ability of macaque monkeys to retrieve, upon instruction, specific features of a unique event (i.e., an episode) in short-term memory in a challenging laboratory task that was originally designed for human subjects (Cadoret et al. 2001) answers in the affirmative the critical question posed by Griffiths et al. (1999), namely whether episodic-like memory in nonhuman animals can be shown to exist outside the confines of certain naturalistic behaviors, such as the food-caching behavior of certain species of birds. The results of the present study demonstrate in an unequivocal manner 2 facts: 1) episodic-like memory retrieval from short-term memory in macaque monkeys which is not an adaptive requirement in the context of a naturalistic behavior, as was the case with the food-caching experiments in the only species of birds (scrub jays) in which such memory has been shown; 2) even more importantly, memory retrieval in macaque monkeys can be the result of explicit active controlled retrieval of specific features of an event in short-term memory triggered by a particular instruction. This characteristic of memory had never before been demonstrated in a nonhuman animal and was thought to be unique to humans. The present results demonstrate that, in the nonhuman primate brain, as in the human brain, memory for events operates at a high abstract and controlled level, that is, it is possible to retrieve specific features from a memory representation upon specific instruction (controlled operation). These new findings add an impressive dimension to the demonstrated nonhuman primate memory capacity and open up new avenues for the study of the neural mechanisms of active controlled retrieval from event memory.
Neuronal Activity in the Ventrolateral Prefrontal Cortex
There is now considerable evidence that the mid-ventrolateral prefrontal cortex in the human brain is involved in active controlled memory retrieval processes (Petrides 2002, 2005; Badre et al. 2005). In earlier functional magnetic resonance imaging (fMRI) experiments with human subjects, we had developed the present paradigm to demonstrate that the mid-ventrolateral prefrontal cortex (area 47/12 and area 45) is a critical region of the human brain for active retrieval processing (Cadoret et al. 2001; Kostopoulos and Petrides 2003). However, fMRI studies in normal human subjects can indicate the engagement of a cortical area in a particular cognitive process by demonstrating changes in its blood oxygenation level (i.e., the blood oxygen level–dependent [BOLD] signal) in relation to that cognitive process, but cannot provide direct information on the neuronal activity that underlies the cognitive process. In the present study, therefore, we recorded neurons in the monkey mid-ventrolateral prefrontal areas 45 and 47/12 that have been shown to be comparable architectonically to those in which the BOLD signal changes were observed in the human brain. The neuronal responses observed in the mid-ventrolateral prefrontal cortex during the performance of the active controlled memory retrieval task were related to the cue period when the retrieval of specific aspects of the encoded event was instructed, the delay that followed the instruction, and the test period when the monkeys had to decide whether the test stimulus matched the specific feature of the memory representation that the cue had instructed.
In the present task, a compound stimulus consisting of an object at a specific location or a shape with a particular color is stored in short-term memory during an encoding event. A cue that follows a little later instructs retrieval of one part of this compound stimulus (e.g., retrieval of location) and the subject must isolate the relevant information (i.e., the location) from the memory representation. One way of thinking about this selective retrieval is to assume that attention is shifted to the appropriate feature of the compound stimulus in memory: attention highlights the location, but not the object, in the memory representation of the compound stimulus, in the example above. The differential cue-related neurons observed in the present experiment (e.g., those that change their firing rate to the object instruction or the location instruction) could reflect the instruction to switch attention to the relevant dimension (Fig. 3B), and, therefore, when the test stimulus appears, it can be matched to the highlighted feature of the memory representation.
Note that, in the present task, in which the subject (monkey or human) is given an instruction (by means of the cue) to engage or not to engage in retrieval across many recurring trials, the nonretrieval of the information is not a passive (by default) occurrence, but rather an active decision not to retrieve. Thus, it was interesting to observe that the activity of some neurons in the mid-ventrolateral prefrontal cortex specifically signaled that no retrieval should be initiated on a particular trial, in contrast to other neurons that signaled that retrieval should be initiated. For instance, the neuron illustrated in Figure 3A responded specifically to the cue indicating no retrieval, whereas the neuron illustrated in Figure 3B did not respond to the cue instructing no retrieval, but responded to the cues instructing retrieval and, in particular, retrieval of the shape feature of the stimulus in memory.
Earlier single-cell recording studies in the prefrontal cortex had reported neuronal activity related to the coding and maintenance of specific stimuli during the delay, the matching and nonmatching decisions inherent in the task, and the rules related to the performance of the tasks (e.g., Fuster 1973, 2000; Funahashi et al. 1989, 1993; Goldman-Rakic 1995; Rainer et al. 1999; White and Wise 1999; Hoshi et al. 2000; Wallis et al. 2001; Goldman-Rakic and Leung 2002; Mansouri et al. 2006). However, these earlier studies could not have examined neuronal activity related to the selective isolation of a specific feature from a multicomponent stimulus in memory because the appropriate paradigms had not been developed. The present paradigm allowed us to demonstrate, for the first time, that the activity of a class of neurons in the mid-ventrolateral prefrontal cortex is related to the process of retrieving a selective feature of a multicomponent stimulus in memory and the cognitive decision that the specific feature has been retrieved. During the delay that followed the instruction cue to retrieve a specific feature of the multicomponent representation in memory, some neurons increased their firing rate selectively after an instruction to retrieve a specific component (Fig. 4B). Furthermore, during the test phase, the activity of another class of neurons, the active controlled retrieval decision neurons (see Figs 5 and 6), was not related to the perception of a specific stimulus, to a specific feature of a stimulus in memory, to the matching/nonmatching (same/different) comparison, or to an instruction cue, but rather to the cognitive decision that the instructed feature of the multicomponent stimulus in memory has been retrieved. Note that the activity of these neurons must reflect the integration of 3 separate events in working memory: a) the instruction conveyed by the cue to retrieve a specific aspect of the memory representation, (b) the isolation of (focusing of attention to) the relevant feature of the multicomponent stimulus in memory, and (c) the relevant feature of the test stimulus (see Fig. 5D and upper part of Fig. 6). The integration in the activity of these neurons of all 3 events relevant to the correct decision, namely, that the instructed feature of the multicomponent stimulus in memory has been retrieved, enables these neurons to code the cognitive decision that the required dimension has or has not been isolated and retrieved.
Note that the activity of these memory decision neurons resolves the potential ambiguity in memory representations that arises when stimuli are related to each other across trials in multiple and more or less equiprobable ways, and there are therefore no strong links between them. For instance, note that, in the present study, some neurons coded the decision that there was a match between “the to-be-retrieved component” (e.g., the object) of the memory representation with the corresponding component of the test stimulus (Fig. 6). As pointed out above, such a neuron can only code this decision by integrating information about the instruction cue, the stimulus in memory and the current test stimulus (upper part of Fig. 6). Thus, the output of these neurons codes the matching and/or nonmatching cognitive decisions based on the integration of the instruction cue, the memory representation, and the test stimulus. In other words, these neurons are ideal for deciding whether object A has or has not appeared in context X on a specific occasion in situations in which several objects (e.g., A, B, C, D) have appeared in certain contexts (e.g., X, Y, Z, W) in multiple and nonpreferential ways. In such a situation, there is an inherent ambiguity in the memory representations leading to increased interference. The memory decision cannot be made automatically (bottom-up), that is, by object A automatically triggering context X, because object A is not uniquely or even strongly linked to context X. A top-down control process that we have called active retrieval is therefore needed to disambiguate in these situations. The cognitive decision neurons in the mid-ventrolateral prefrontal cortex discovered in the present investigation can make this decision by integrating information about the instruction (which leads to the intention) to retrieve a certain piece of information, the memory representation of the encoded event and the information that has just entered short-term memory as a result of the presence of the test stimulus (see Fig. 5D and upper parts of Figs 6 and 7).
This decision that underlies active memory retrieval processing probably requires an interaction between mid-ventrolateral prefrontal cortex and specific posterior association cortical areas in which memory representations are thought to be stored. In a study with split-brain monkeys, Tomita et al. (1999) provided evidence that the prefrontal cortex can influence visual object memory retrieval from the inferior temporal cortex. This evidence is consistent with the hypothesis that, in the absence of bottom-up sensory inputs, top-down influences originating from the prefrontal cortex can regulate memory retrieval (Cadoret et al. 2001; Kostopoulos and Petrides 2003; Miyashita 2004). Indeed, it has been argued that the bottom-up signals account for automatic memory retrieval, whereas the top-down signals account for active controlled memory retrieval in the paradigm studied here (Cadoret et al. 2001).
The neuronal response associated with the disambiguation reported here is also consistent with the results of lesion studies of the ventral part of the lateral prefrontal cortex. Dias et al. (1996) have shown that, when faced with a series of compound discriminations (i.e., stimuli that are composed of 2 dimensions, as in the present experiment), monkeys with ventral lateral prefrontal lesions are not impaired in learning to discriminate between different exemplars of compound stimuli, provided the relevant dimension (e.g., shape) remains the same across trials (intradimensional shifts of attention). These monkeys, however, are impaired in learning to discriminate between exemplars of compound stimuli when they are required to shift attentional set, that is, select between the stimuli on the basis of another dimension (extradimensional shifts of attention). The “disambiguation” neuron reported in the present experiment can be thought of as a correlate of the shifting of attentional control from one dimension to another from trial-to-trial in the context of a short-term memory task. The neuronal response reflects the relating of a specific dimension of a compound stimulus that is currently in short-term memory with the relevant dimension of a compound stimulus on the screen (test stimulus) based on an instructional cue that shifts attention to one or the other dimension of the stimuli.
In conclusion, the activity properties of the various neurons in the mid-ventrolateral prefrontal cortex observed in the present study provide the basis for a neuronal mechanism enabling the isolation of the relevant feature (information) from memory representations and the disambiguation of representations in memory when stimuli enter in multiple relations with other stimuli/contexts across trials, and, therefore, ambiguity (interference) is generated. The mid-ventrolateral prefrontal cortex contains neurons instructing retrieval of specific information from a multicomponent memory representation (Fig. 3), neurons reflecting specific aspects of retrieval during the postinstruction delay period in anticipation of the test phase (Fig. 4), and neurons that, during the test phase, integrate information from the instruction, the stimulus in memory and the test stimulus on the screen in order to code the cognitive decision that the relevant information has been retrieved (Figs 5, 6, and 7). These single-neuron recording findings, therefore, provide an initial glimpse into the neuronal properties of a prefrontal control module, the mid-ventrolateral prefrontal region, which in interaction with relevant posterior cortical and subcortical regions, underlies active controlled retrieval from memory (Petrides 2005).
Medical Research Council of Canada fellowship to G.C.; Natural Sciences and Engineering Research Council of Canada (Individual Research Grant 7466), and the James McDonnell Foundation to M.P.
We thank Mireille Bouchard for excellent technical assistance, Philippe Drapeau for programing, and Veronika Zlatkina and Emily Rubin-Ferreira for preparing Figure 2. Conflict of Interest: None declared.