Item to Decision Mapping in Rapid Response Learning

Abstract

Repeated classification of a visually presented stimulus rapidly leads to a form of response learning that bypasses the original evaluation in favor of a more efficient response mechanism. Two experiments examined the level of input and output representations that make up this form of learning. In Experiment 1, alterations in the finger mapping of the output response had no effect on the expression of response learning, demonstrating that a classification decision, not motor output, is associated with repeated items. Experiment 2 tested whether response learning transferred across different visual exemplars of a studied item. There was no evidence of transfer to different visual exemplars, even when these exemplars were judged to be highly visually similar. Taken together, these results indicate that response learning consists of the formation of an association between a specific visual representation and a classification decision.

A previous encounter with an item will often result in changes in a person’s ability to identify, produce or classify that item, referred to as repetition priming. Simple changes in presentation format or task demands between exposures can have significant effects on the magnitude of priming (Schacter, Dobbins, & Schnyer, 2004). The nature and type of these specificity effects may be critical in helping to identify the representational level, or levels, upon which behavioral facilitation rests. For example, if a repetition priming effect were severely disrupted by changing the nature of the manual response, or by changing from a manual to verbal response, this finding would suggest that the learning or facilitation occurred at a relatively late stage of processing. Despite the known sensitivity of priming to a variety of test manipulations (Burgund & Marsolek, 1997; Vaidya, Gabrieli, Verfaellie, Fleischman, & Askari, 1998) there have been relatively few attempts to delineate the mechanisms that are responsible for these effects. In a recent series of studies (Dobbins, Schnyer, Verfaellie, & Schacter, 2004; Schnyer, Dobbin, Nicholls, Schacter, & Verfaellie, 2006) we have begun to explore the cognitive and neural mechanisms resulting in one type of priming specificity that suggests that the rapid learning of decision outcomes or responses may be a significant part of typically observed priming gains. We have referred to this phenomenon as response learning.

Previous Examinations of Response Learning

In previous studies, we have examined response learning utilizing a semantic classification task in which the framing of the decision cue was changed between initial exposure (study) and subsequent primed presentations (test). During a study period, participants were asked to indicate whether visually presented common objects were “bigger than a shoebox”. Items were presented during study either once or 3 times. At test, participants continued making size decisions to objects repeated from study and new objects. The test phase was conducted with either the same decision cue as during study (“bigger than a shoebox”) or with an inversion of the decision cue (“smaller than a shoebox”). The effect of repetition on items presented once or three times prior was compared between the 2 decision cue conditions. We postulated that if the facilitation associated with repetition reflected small modifications in object identification and/or knowledge representations within the same object processing stream engaged when items are first presented, then inversion of the decision cue would cause little disruption in priming. That is, if priming were the result of learning or tuning in these representations (Wiggs & Martin, 1998), then its expression should not be particularly sensitive to the change in decision cue. However, if the results demonstrated that priming was significantly disrupted or eliminated by a change in the decision cue, then this would suggest that priming reflects the fact that participants rapidly associate their prior responses or decisions with each item, and thereby bypass many of the more deliberative processes engaged when an item is first presented.

Results provided evidence for the latter view (Dobbins et al., 2004): cue inversion disrupted a significant portion of the response facilitation associated with repetition. Moreover, as indicated by fMRI, cue inversion completely eliminated the neural priming signature across several cortical areas where repetition-related reductions in neural activity were evident, including the fusiform gyrus and left prefrontal regions. In addition, regression analysis demonstrated that changes in prefrontal cortex predicted the level of behavioral facilitation during learning, and the magnitude of behavioral disruption that would later occur when the cue was inverted. Given this outcome, we tentatively suggested that the behavioral facilitation and neural activity reductions, both in regions associated with early visual processing and in those associated with later classification and decision making processes, resulted from the learning of prior responses or decisions that enabled observers to bypass controlled or deliberative classification processes that would be engaged when an item is first encountered. This view differs significantly from one offered in previous fMRI studies of repetition priming (Buckner et al., 1998; Koutstaal et al., 2001), postulating that the facilitation resulting from multiple repetitions reflects continued “tuning” of object identification and knowledge systems.

In a subsequent set of behavioral experiments, Schnyer, et al. (2006) replicated the finding that the level of priming was reduced with cue inversion from that seen when the cue orientation was maintained from study to test. Unlike the fMRI study where cue sensitivity was apparent primarily for items repeated three times, here cue inversion disrupted priming for items repeated once as well as those repeated three times. Moreover, in this experiment it was apparent that all the added facilitation associated with multiple repetitions was eliminated by cue inversion. In a second experiment, the effect of cue inversion was tested in amnesic patients with damage to the medial temporal lobe (MTL).These patients revealed a significant behavioral priming effect but no evidence of a response learning component. Unlike controls, MTL amnesics did not reveal added facilitation with multiple repetitions and the observed priming was not disrupted by cue inversion. This study led to the conclusion that rapid response or decision learning reflects an MTL dependent associative learning mechanism whereby a particular response or decision outcome becomes associated with an item.

It is important to note, however, that the presence of cue specificity effects, as documented above, does not imply that all of the behavioral facilitation observed in classification priming is dependent upon the same mechanism. Indeed, although a considerable portion of the behavioral facilitation was dependent upon the format of the decision cue, there was nonetheless a portion of repetition-induced facilitation that appeared independent of the specific format of the cue and that was not impaired in amnesia. These findings suggest that priming, even within a single task, may reflect multiple mechanisms or processes, only one of which reflects the learning of prior decisions or responses. In this way, our results and conclusions differ from previous demonstrations of response learning that emphasize a common underlying mechanism of “instance learning” that is solely responsible for repetition priming (Logan, 1990; Logan & Etherton, 1994; Logan, Taylor, & Etherton, 1996).

Delineating the Input/Output Levels Bound in Response Learning: The Present Experiments

The current experiments are focused on pinning down the input and output levels of the bound association that is formed in rapid response learning. On the output side, the question arises as to whether the effect reflects the learning of an association between an item and a specific motor/finger response, or rather, between an item and a specific classification/decision. So far the evidence from previous studies is mixed. In many of the previous studies of instance learning, motor mapping is assumed to play a minimal role (Dennis & Schmidt, 2003; Fisk & Schneider, 1984; Logan, 1990). Instead, it has been assumed that instance learning reflects the formation of an association between items and response categories, or classification actions (Logan, 1990). More recently, however, it has been demonstrated that conditions can be implemented that result in interactions between motor mapping and other features of the learned association (Logan et al., 1996).

The role of motor mapping in the object classification task we have utilized in our previous studies is also unclear. In Dobbins et al. (2004), subjects were tested with a version of the task outside of the MRI scanner that left the decision cue intact (“bigger than a shoebox”) but reversed the assignment of “yes” and “no” response buttons across study and test. In that experiment, a sizeable priming disruption occurred, but more importantly, there was no statistical interaction between the effect of decision cue inversion and the effect of motor mapping inversion. This result suggested that the form of response learning exhibited in the object classification task is predominantly driven by motor associations and conflicted with some of the previous studies mentioned above.

There are a number of reasons to be skeptical of our previous motor mapping results. First, the experimental design utilized in that study (which was adapted to parallel the one used during the fMRI experiment) contained several suboptimal methodological features. The critical comparison between the same and the inverted cue condition suffered from a temporal confound, in that the inverted cue condition was tested later in the experiment than the same cue condition. In addition, items in the inverted cue condition had received one extra repetition in comparison to those in the original cue condition. Any difference between study and test blocks could be driven in part by changes in response speed with time or the additional repetition. Given this ambiguity, it is not possible to draw any firm conclusions about whether the output level in the classification task is bound to the motor output or the decision. Experiment 1 was designed to address this issue.

The goal of Experiment 2 was to determine the input level of response learning by examining the perceptual specificity of the effect, given evidence that priming is often perceptually highly specific (Koutstaal et al., 2001). If the response learning phenomenon appears tied to highly specific perceptual representations, then it is possible that disruptions in priming across perceptual manipulations in earlier research (Schacter et al., 2004) may in some cases have been due to disruptions of response learning. While a number of studies have examined object priming across different visual exemplars (Biederman & Gerhardstein, 1993; Koutstaal et al., 2001; Roediger & Srinivas, 1993), there are no studies which have tested whether the priming preserved across object exemplars is, or is not, sensitive to the format of the decision cue. Thus the level of perceptual specificity of the response learning mechanism is unknown.

EXPERIMENT 1

In Experiment 1, we evaluated the effect of changes in motor mapping on priming in the classification paradigm used in our previous work (Schnyer, et al. 2006), by introducing a switch in the finger mapping of the “yes” and “no” response keys. If a simple change in finger assignment results in a significant reduction in priming, as does cue inversion, this finding would support the view that during response learning, items become associated with specific motor operations. Alternatively, if a switch in finger mapping has no significant effect on response learning, it would suggest that during response learning, items become associated with a specific classification (e.g. “bigger than”).

Method

In order to compare directly the results of this experiment with our previous examination of response learning, the stimuli, paradigm and procedure were exactly the same as in Schnyer, et al. (2006). The only difference was that rather than inversion of the decision cue, subjects were asked to invert which finger they used to indicate the “yes” and “no” responses.

Participants

Sixteen young native speakers of English (2 male, 14 female), with normal or corrected to normal vision took part in the experiment. Participants (mean age = 21 years, range 18-25) were recruited through flyers and advertisements at local colleges and universities and received $10 for their participation. Participants were screened using a short medical questionnaire to ensure that they were free from current psychiatric or neurological disorder, any history of brain injury or excessive drug or alcohol use. Written informed consent was obtained from each participant prior to the experimental session. The Human Subjects Committees of Boston University School of Medicine and the Veterans Affairs Healthcare System approved all procedures.

Stimuli

Four hundred and eight colored line drawings of common animate and inanimate objects were selected from commercially available clip-art collections (CDROM from Corel Mega Gallery, Corel Corporation, 1997). Pictures reflected varying orientations and visual size. The stimuli were presented on a Mac Powerbook laptop computer using Psyscope 1.2.5 (Carnegie Mellon University, 1994). Objects were presented within a centrally located 8.75 x 8.75 cm box and viewing was approximately 75 cm from the screen, resulting in a vertical and horizontal visual angle subtending approximately 6-7 degrees.

Procedure

A brief task instruction period was followed by 4 alternating ‘study-test’ cycles. During the study phase of each cycle, 34 pictures were presented once and 34 pictures were presented three times, for a total of 136 presentations. Items presented only once were evenly distributed throughout the study phase in such a way that one third of these items were encountered with each full repetition cycle of the items presented three times total. Participants were asked to make a size judgment by deciding if in real life the depicted object was “bigger than a shoe box”. They indicated their decision by pushing a “yes” or a “no” key with the index and middle fingers, respectively, of their right hand. Following the study phase and a short pause, participants took part in a test phase consisting of two test blocks. Each test block consisted of 17 pictures presented once during study (low prime), 17 pictures presented three times during study (high prime), along with 17 novel pictures. None of these pictures were repeated within or between test blocks. The test blocks differed only with regards to the finger mapping of the “yes” and “no” decisions made by participants: In one block participants indicated their response with the same mapping (<yes/no>) made during the study phase, whereas in the other block the mapping was inverted (<no/yes>) with respect to the finger that indicated each response. For two study/test cycles, participants used the same mapping in the first test block as they had used in the study phase and the inverted mapping in the second test block, whereas for the other two cycles the order was switched.

Pictures were randomly assigned to one of the 4 study-test cycles. Within each cycle, pictures in the test phase were rotated among the three possible conditions (novel, low prime, high prime). Additionally, pictures in the test phase were rotated so that pictures occurring in the same cue condition for one participant occurred in the inverted cue condition for another. This resulted in a total of 6 versions of the experiment.

Pictures were presented at a rate of 1 every 2 seconds and were accompanied at the bottom of the screen by the finger mapping to be used to indicate the size decision on that trial. Instructions appearing with the stimuli were “Bigger than a shoebox, <yes/no>” during the study phase and same mapping at test and “Bigger than a shoebox, <no/yes>” when the finger mapping was inverted at test. Examples of the pictures and the basic paradigm can be seen in Figure 1.

Figure.1.

The general experimental paradigm. Participants engaged in a study period where they made classification judgments about items presented once or three times. At test items presented once (low prime) and thrice at study (high prime) were presented along with novel items. Half of the time at study participants responded with the same finger mapping as during study and half of the time with an inverted mapping.

Results

Response consistency:

In order to determine if participants were switching responses appropriately, the consistency of their responses was measured by comparing the last response given to an item at study and the response indicated at test. If the response key mapping remained the same then subjects should maintain the same response. In contrast, if the response mapping changed, then in order to remain consistent, subjects would have to change the keyed response. Overall, the group was highly consistent (93%, SD = 4.1%). However, a single subject was more than 2 standard deviations below the mean (77%) and that person’s data was eliminated from the analysis. Because the origin of inconsistent responses is unclear, these responses were also eliminated (6% of items overall).

Study phase priming:

Response times during the study phase demonstrated facilitation across repetitions. An ANOVA examining RTs with presentation (once, twice, thrice) as the within-subjects variable revealed a significant main effect (F(2,28) = 202.13, p < .001), and pair wise t-tests indicated that all conditions were significantly different from one another (mean RT first presentation = 1127 ms, second presentation = 947 ms and third presentation = 884 ms, all p < .001).

Finger mapping inversion:

Reaction times across all test block conditions can be seen in Table 1. The effect of inverting the finger mapping on priming was examined in test block 1, where the contrast between same and inverted mapping is not confounded by number of repetitions or temporal order. Proportional priming scores were calculated ([novel-repeat]/novel) and submitted to a 2 X 2 ANOVA with finger mapping (same, inverted) and priming condition (low prime, high prime) as within-subject factors. There was a main effect of priming condition (F(1,14) = 39.31, p < .001), which reflected the fact that priming in the high prime condition was greater than that in the low prime condition. Neither the main effect of finger mapping (F(1,14) = 1.80, p > .20), nor the finger mapping by priming condition interaction was significant (F(1,14) < 1), indicating that finger mapping inversion had no effect on the level of priming for either high or low primes. Follow-up t-testing revealed that priming in all conditions was significantly greater than zero (all ps < .001, Figure 2 - block 1 same mapping and block 1 inverted mapping).

Table 1.

Mean response times and standard errors for Experiment 1. Values are in milliseconds.

	BLOCK 1			BLOCK 2
Finger Mapping	Novel	Low	High	Novel	Low	High
SAME
mean RT	1078	966	865	1055	928	890
(SE)	(56)	(65)	(52)	(64)	(59)	(50)
INVERTED
mean RT	1054	960	873	1145	1035	963
(SE)	(56)	(59)	(56)	(61)	(59)	(60)

Figure.2.

Proportional priming scores for low and high primed conditions across decision cue inversion in test blocks 1 and 2. Graphs represent percent priming and error bars represent standard error of the mean. The y axis is in units of percentage.

In order to determine if finger mapping inversion disrupted response learning after a longer period with the original mapping, we compared priming scores when the finger mapping reversal occurred during the first test block to when it was delayed until the second test block. A 2 X 2 repeated measures ANOVA with block (1,2) and priming condition (low prime, high prime) as factors again resulted in only a main effect of priming condition (F(1,14) = 35.60, p < .001), continuing to reflect the overall difference in priming scores for high and low primed items. Again there was no evidence of an effect of finger mapping inversion, as evidenced by the fact that neither the main effect of mapping (F(1,14) < 1) nor any interaction involving mapping (F(1,14) < 1) were significant. This result indicated that there was no effect of finger mapping inversion even after a longer period of responding with the original finger mapping assignment (Figure 2 - block 1 inverted mapping and block 2 inverted mapping).

Discussion

The results of Experiment 1 indicate that a change in the motor mapping of the response keys had no discernable effect on priming, either following a single presentation or following multiple presentations. Moreover, there was no effect of finger mapping inversion, regardless of whether the change in decision cue occurred in test block 1 or in test block 2. Thus having eliminated the confounds that were present in the previous study (Dobbins, et. al, 2004), our results are now in agreement with those obtained in other labs (Dennis & Schmidt, 2003; Logan, 1990) and add to the evidence that the critical associative link that is forged during response learning is between a stimulus and its associated decision rather than its motor response (Logan, 1990).

Having clarified the output level that is bound in response learning, we now turn to an examination of the input level that is bound to the learned classification. It is possible that this binding occurs at the level of relatively abstract visual or even semantic representations; alternatively, the binding may be specific to the particular visual form that is presented during the study phase. One way to address this question is to evaluate response learning when different visual exemplars are presented at study and at test. Previous studies have shown that presenting different visual instantiations of the same object across repetitions reduces the level of priming relative to when the same exemplar is repeated (Biederman & Gerhardstein, 1993; Roediger & Srinivas, 1993) and functional imaging studies have demonstrated that similar visual form manipulations reduce both behavioral and “neural priming” effects in early visual processing areas (Koutstaal et al., 2001; Simons, Koutstaal, Prince, Wagner, & Schacter, 2003). However, it is important to note that in these studies facilitation was not completely eliminated by substituting exemplars across repetitions. That is, although priming was significantly reduced, subjects still responded significantly faster to these alternate exemplars than they did to completely new items. Whether this residual priming represents response learning is unknown.

If response learning reflects an association between a relatively abstract visual or semantic representation and a classification response, then it should be preserved across visual form changes in repeated objects. Alternatively, if response learning reflects an association between a specific visual form and a classification response, then response learning should be disrupted by changing the visual exemplar across repetitions. Experiment 2 examines the effect of cue inversion on same- and different-exemplar repetitions in order to distinguish between these two possibilities.

EXPERIMENT 2A

Participants

Seventeen young native speakers of English (4 male, 13 female), with normal or corrected to normal vision took part in the experiment. None of these individuals had participated in Experiment 1. Participants (mean age = 20 years, range 18-23) were recruited through flyers and advertisements at local colleges and universities and received $10 for their participation. Participants were screened as described in Experiment 1. Written informed consent was obtained from each participant prior to the experimental session. The Human Subjects Committees of Boston University School of Medicine and the Veterans Affairs Healthcare System approved all procedures.

Stimuli and procedure

Three hundred items were selected for the current experiment from the items used in Experiment 1. Each of the selected items had a matching exemplar that differed visually but elicited the same name. The level of name correspondence between exemplars was previously established behaviorally (Koutstaal et al., 2001). The items were divided into 4 groups and item counterbalancing was accomplished by rotating the 300 items through the 4 conditions of once repeated, thrice repeated same exemplar, thrice repeated different exemplar, and novel. Objects were presented centrally and viewing was approximately 75 cm from the screen. The experiment was conducted on a PC notebook computer running DMDX (software developed at Monash University and at the University of Arizona by K.I.Forster and J.C.Forster.).

The procedure was modified from Experiment 1 in two important ways. First, since previous studies have indicated that the most robust level of response learning occurs with 3 repetitions (Dobbins et al., 2004; Schnyer et al., 2006), the current experiment carried over into the test phase only items presented 3 times during study. Single items were included during the study phase so that the study conditions corresponded to those used in our previous experiments (Experiment 1 and Schnyer, et al., 2006) and to ensure that not only repeated items were appearing by the end of the list. Additionally, in order to simplify the analysis, each test phase was restricted to a single block with either cue inversion or same cue conditions. Participants engaged in 6 ‘study-test’ cycles, with the test phases alternating between cue inversion and same cue. The order in which the two cue conditions were presented was counterbalanced across subjects.

During each study phase, participants made size judgments to 80 items total, consisting of 20 items repeated 3 times and 20 items presented once. Items were presented in pseudo random order, with items presented once evenly distributed throughout the study phase. None of the items presented once during study were seen during the test phase. Participants were asked to make a size judgment by deciding if the real life object depicted in the picture was “bigger than a shoe box” and to indicate their decision by pushing a “yes” or a “no” key with the index and middle fingers of their right hand, respectively.

Each test phase consisted of 30 randomly ordered pictures. Ten of these were exact copies of objects presented 3 times during study (same exemplar), 10 were visually different exemplars of objects presented 3 times during study, (different exemplar) and 10 were objects not seen during study (novel). The division between the study and test phase was indicated by a screen that instructed the participants whether to continue with the same size decision or whether to invert that decision (“smaller than a shoe box”).

Pictures were presented at a rate of 1 per 2 seconds and were accompanied at the bottom of the screen by the appropriate decision cue. Instructions appearing with the stimuli were “Bigger than a shoebox, <yes/no>” during the study phase and the same-cue test condition and “Smaller than a shoebox, <no/yes>” during the inverted-cue test condition.

Results

Response consistency:

Similar to Experiment 1, participants were highly consistent in their responses (94%, SD = 2.9%), as evidenced by nearly identical classification of items during the third study phase presentation and the test phase presentation. A single subject performed more than 2 standard deviations below the group mean (86%) and that person’s data were eliminated from the analysis. For all analyses of RTs, inconsistent responses were eliminated (5% of items overall).

Study phase priming:

As in our previous experiments, increasing repetition resulted in increased RT facilitation. An ANOVA examining RTs with presentation (once, twice, thrice) as the within-subjects variable revealed a significant main effect (F(2,30) = 176.55, p < .001), and post-hoc testing indicated that all conditions were significantly different from one another (mean RT first presentation = 949 ms, second presentation = 799 ms and third presentation = 739 ms, all p < .001).

Decision cue inversion:

The proportional priming scores across cue orientations and exemplar status are shown in Figure 3 and the reaction times for all conditions can be seen in Table 2. Priming in all conditions was significantly greater than zero (all ps < .005). Proportional priming scores were examined in a 2 X 2 repeated measures ANOVA with decision cue (same, inverted) and exemplar condition (same exemplar repeated, different exemplar repeated) as within-subject factors. There was no overall effect of decision cue (F(1,15) = 2.04, p > .15), while there was an effect of exemplar condition (F(1,15) = 38.30, p < .001), which reflected the fact that priming was greater for repetition of the same visual exemplar compared to repetition of different visual exemplars. Most importantly, there was a significant decision cue by exemplar condition interaction (F(1,15) = 5.03, p < .05). For the same exemplar repetitions, priming was greater in the same than in the inverted cue condition (t(15) = 2.46, p < .05). By contrast, for different exemplar repetitions, there was no difference in priming in the same and inverted cue condition (t(15) < 1). Thus, there was no evidence that the portion of priming sensitive to decision cue inversion (and indicative of response learning) transfers across different visual exemplars of a studied item.

Figure.3.

Proportional priming scores for same and different visual exemplars across both decision cues. Graphs represent percent priming and error bars represent standard error of the mean. The y axis is in units of percentage.

Table 2.

Mean response times and standard errors for Experiment 2A. Values are in milliseconds

Decision Cue	Novel	Same Exemplar Primed	Different Exemplar Primed
SAME
mean RT	901	730	831
(SE)	(31)	(27)	(30)
INVERTED
mean RT	969	854	888
(SE)	(37)	(33)	(28)

EXPERIMENT 2b

To determine whether the transfer of response learning across exemplars depends on the level of visual similarity, a separate group of subjects made visual “similarity” judgments for the exemplar pairs used in Experiment 2. These judgments were then used to sort the items based on visual similarity in order to determine if sensitivity to cue inversion dependent on the visual similarity of exemplars.

Participants

Twelve young adults (10 female, 2 male), average age 23.6 years (range 18-32) participated in the similarity ratings. None of these individuals had participated in Experiment 2. Participants were recruited through flyers and advertisements at local colleges and universities and received $10 for their participation. Participants were screened as described in Experiment 1 and 2. Written informed consent was obtained from each participant prior to the experimental session. The Human Subjects Committees of Boston University School of Medicine and the Veterans Affairs Healthcare System approved all procedures.

Stimuli and procedure

Each of the 12 participants saw all 300 image pairs that were utilized in Experiment 2 across 3 separate runs of 100 pairs each. Pairs of images were shown side by side at the center of the screen and participants had a total of 20 seconds in which to rate how visually similar they thought the images to be. Once the rating was given for a particular item, the next pair was presented. Before beginning, participants were shown a few example pairs and were given these instructions:

In this task you will see two images side by side and will be asked to rate how visually similar they are. There is no right or wrong answer. These images were used in a previous study and we just want to know how visually similar people think they are. You can base your similarity judgments on the shape, orientation and design of the images. However, please do not use the color or size of the images to make your similarity judgments. You will make your rating on a scale from 1 to 6 using this scale: 1 - extremely dissimilar, 2 - dissimilar, 3 - slightly dissimilar, 4 - slightly similar, 5 - similar, 6 - extremely similar.

Results

Mean similarity ratings were generated for each of the 300 pairs of items. These mean ratings were then used to sort the different exemplar conditions from Experiment 2A into 5 bins - bin1 = items rated 1-1.99 (2%), bin2 = 2-2.99 (24%), bin3 = 3-3.99 (41%), bin4 = 4-4.99 (81%), bin5 = 5-6 (6%). Bin1 was dropped from the analysis because there were too few items. The reaction times from experiment 2A sorted by similarity rating for the same and inverted cue conditions can be seen in Table 3. Mean proportional priming scores from Experiment 2A were re-analyzed in a 2 X 4 repeated measures ANOVA with decision cue (same, inverted) and level of judged similarity (bin2, bin3, bin4, bin5) as within subjects factors. There was no effect of decision cue orientation (F(1,11) < 1) but there was a significant effect of similarity rating (F(3,33) = 3.57, p < .01). Planned Post hoc comparisons indicated that only the low similarity rating bin2 revealed less priming relative to all the other bins (p < .05) and priming was not significant for that bin (t(15) < 1, see Figure 4). Importantly, there was no evidence that the similarity rating interacted with the decision cue orientation (F(3,33) < 1). Thus, priming was unaffected by inversion of the decision cue, regardless of the visual similarity of exemplars, suggesting that response learning did not transfer across any exemplars. A graph of the level of response facilitation across the level of visual similarity for both same and inverted response cues, along with examples of the pictures and their corresponding exemplar for bin 2 and bin 5 can be seen in Figure 4.

Table 3.

Mean response times and standard errors from Experiment 2A sorted across the 4 similarity rating bins generated in Experiment 2B. Values are in milliseconds.

	Similarity Rating Bin
Decision Cue	5-6	4-5	3-4	2-3
SAME
mean RT	826	821	838	911
(SE)	(67)	(34)	(41)	(38)
INVERTED
mean RT	864	879	905	962
(SE)	(86)	(44)	(44)	(30)

Figure.4.

A graph of the proportional priming scores across levels of visual similarity between exemplars for both the same decision cue and the inverted decision cue. From left to right, values reflect repetitions of the same visual exemplar followed by the visual similarity rating bins 5, 4, 3, and 2 for repetitions of different visual exemplars. Located above the graph are representative examples of the visual objects at the 2 extremes of the similarity ratings. Markers represent percent priming and error bars represent standard error of the mean. The y axis is in units of percentage.

Discussion

Experiment 2 demonstrates that response learning in a classification task does not transfer across different visual exemplars of the same object, even when the perceptual changes are highly subtle. This is evident from the fact that decision cue inversion had no effect on the magnitude of priming for different exemplars. Furthermore, even when different exemplars were analyzed based on subjective ratings of visual similarity, there was still no evidence of response learning, as evident from a lack of any effects of decision cue inversion even for the exemplars judged to be most highly similar. Priming was observed across exemplars, however, and similar to a previous study using multiple study repetitions (Koutstaal et al., 2001), the magnitude of priming for different exemplars was significantly smaller than that exhibited for repeated same exemplars. The only exception to this was items judged to be very dissimilar; these items evidenced no priming at all. It is difficult to interpret why there is no priming across dissimilar exemplars but one possibility is that these items have lower name agreement, and as such, a significant portion of them may be considered fundamentally different items.

The current results provide the first evidence that response learning in an object classification task is perceptually highly specific and is not preserved across visually similar exemplars. It remains unknown whether there are other object-related perceptual parameters across which response learning might transfer. For instance, it is possible that response learning would not be disrupted by color changes (Cave, Bost, & Cobb, 1996), changes in relative visual size (Biederman & Gerhardstein, 1993) or depth rotation/mirror imaging (Lawson, 2004). In the verbal domain, Logan, Taylor and Etherton (1996) have examined the impact of color changes on classification learning. They found that classification learning for words was not disrupted by a change in the color in which the words were presented, unless subjects were required to attend to the word color at study. Thus, the effect of featural changes on response learning may depend on the relevance of the featural change to the task. However, the correspondence between our work examining classification of visually presented objects and word stimuli (Logan et al., 1996) is not yet well established and important differences may yet emerge.

Also relevant are findings of a study of instance learning that examined how repetition and visual similarity affect the speed of numerosity judgments to random dot patterns (Palmeri, 1997). In that study, subjects saw dot patterns repeatedly throughout a training phase, and in a transfer phase were presented with dot patterns that varied in visual similarity to those in the training phase. As expected, reaction times decreased across the training phase and response latencies during the transfer phase depended on the degree of visual similarity to the original patterns. The presence of instance learning was revealed by an examination of response latencies as a function of the number of dots in each pattern. Whereas early in learning, latencies increased roughly linearly with the number of dots, at the end of training response latencies were independent of numerosity, suggesting that subjects had learned a specific response to each pattern. For transfer patterns with moderate similarity, the slope of the function relating RT to numerosity fell in between that for new and repeatedly trained patterns. This finding might be taken as evidence that for some dot patterns, response learning was preserved despite some degree of visual change, whereas for other patterns a counting algorithm needs to be applied anew. An alternative possibility is that for each pattern, performance reflects some combination of instance learning and application of the counting algorithm. For instance, a subset of dots within the altered array may retain the same visual pattern as in the learning phase. This may allow a participant to rapidly retrieve a previous count for that cluster and then to continue counting the remaining dots. Importantly, by this view, instance learning would only occur for any cluster that does not undergo visual distortion. Regardless of the best account of Palmeri’s (1997) findings, it is important to keep in mind that learning transfer was directly related to the degree of visual distortion, suggesting that instance learning is indeed disrupted by visual change.

The finding in the present study that response learning failed to transfer between different object exemplars, even when judged to have a high degree of visual similarity, can be understood in the context of a theoretical framework that postulates dissociable object representation subsystems for specific and abstract object codes. Previous work utilizing divided visual field presentations (Marsolek, 2004; Marsolek & Burgund, 2003) has demonstrated that the right hemisphere shows greater priming for same than for different exemplars, while the left hemisphere shows equal priming for both. These findings have been interpreted as evidence for the existence of a specific object representation subsystem within the right hemisphere and an abstract object representation system within the left hemisphere (for related ideas, see Schacter, 1994). Further evidence for this notion comes from imaging studies, which have demonstrated that the right fusiform gyrus is differentially sensitive to repetition of specific exemplars (Koutstaal et al., 2001; Simons et al., 2003), whereas the left fusiform gyrus exhibits repetition-related changes more generally across exemplars. Within this framework, the findings of the current study suggest that the association between a response decision and object representation is established within a right hemisphere specific representation system. Future neuroimaging and/or divided field studies will be needed to confirm this neuroanatomical suggestion.

In the current study, all but the highly dissimilar exemplars showed priming, and the magnitude of priming was not directly tied to the nature of the decision cue. In addition, for same exemplars as well, there remained a level of priming that was not disrupted by decision cue inversion. These two findings point to a portion of the facilitation that operates on relatively abstract representations (Bowers, 2000) and this level of representation appears not to be linked directly to outcome decisions. The distinction between abstract representations, which mediate priming across exemplars, and specific representations, which mediate response learning, is adaptive, since each representational system allows for a different level of learning. In the case of the abstract system, generalization across exemplars enables establishment of a general organizational structure that allows for flexible application of novel classification and decision schemes. In contrast, when a specific item is repeatedly associated with a specific decision outcome, there are considerable computational savings associated with the rapid learning of these instances, and this is precisely what appears to be occurring in response learning.

Finally, while there is a well-formulated theoretical framework regarding abstract and specific representations and their influence on priming, we have reached the conclusions from our current work primarily though the presence of statistical independence between components. Because the empirical measure being used (RT) has undergone a number of transformations from the proposed underlying cognitive/neural components, any statistical differences could reflect variance due to these transformations rather than independence of the components. Future converging evidence, neural dissociations and well formulated computational models will be needed in order to provide further support for the proposed framework.

General Discussion

The current experiments suggest that multiple components contribute to priming, and in so doing, present a challenge for several existing theoretical frameworks that assume that a single mechanism accounts for all priming effects. A significant portion of the facilitation associated with the repetition of objects appears to involve the reliance on learned classification responses that bypass processes engaged during the initial classification. These learned responses do not reflect changes in object knowledge systems, and therefore cannot easily be accommodated within abstractionist frameworks (Bowers, 2000) that argue for facilitation or tuning within object knowledge systems (Wiggs & Martin, 1998). Instead, this component is consistent with an instance learning account (Logan, 1990), which postulates that items come to be associated with a learned classification. In this regard, what we have previously referred to as “response learning” can perhaps be more accurately termed “decision learning”.

In addition to decision learning, there appears to be another component of repetition priming that operates across a certain level of visual dissimilarity, and this component is impervious to decision cue inversion. This component demonstrates features that are easily accommodated by an abstractionist framework (Bowers, 2000). By this framework, item repetition induces changes in abstract representations that are not tied to a specific encounter. This level of representation is general enough to allow facilitation to transfer across different visual presentations of the same item.

A question arises as to whether priming across different exemplars, which we take to reflect activation of abstract representations, might also be accommodated by an instance learning framework (Tenpenny, 1995). For example, could it be that priming for different exemplars reflects binding of the decision to a lexical, rather than visual, representation (Logan et al., 1996)? The majority of the visually presented items used in the current study are relatively easy to identify and therefore little attention is required to identify them. However, a small subset of items may be more difficult to identify, and therefore additional attentional resources may be needed to name the object before the classification decision can be made (e.g. “is this item a nail or a broom stick?”). It could be argued that the additional effort associated with naming the item may result in a learned instance that reflects the binding between the name and the decision outcome. Such instances would transfer across exemplars and thus might provide a mechanism for different exemplar priming. Because bound instances at the lexical level would occur less frequently than bound instances at the visual level, the resulting facilitation would be less for different exemplars than for same exemplars.

Inconsistent with this account, however, is the fact that decision cue inversion had no effect on the level facilitation for different exemplar items. If priming for these items reflected a name-decision linkage, then one would expect an effect of cue inversion because cue inversion disrupts the output level of the learned instance. Such disruption should occur regardless of whether the decision is bound to a specific visual representation or a lexical one. Thus, while possible to formulate different associations that could account for these results, it is more likely that priming for different exemplars, both prior to and after cue inversion, is not tied to specific instances. Priming for different exemplars, as well as the component of same exemplar priming that is unaffected by cue inversion, is more easily accommodated in the context of an abstractionist framework. In this regard, our view diverges from one that postulates that repetition priming and the automaticity that emerges from multiple repetitions stem from a unitary mechanism (Logan, 1990).

An unresolved issue with regards to the two components of priming evident in the current experiments is whether they are independent or interactive. One possibility is that they reflect independent parallel pathways to decision output. In this case, the level of facilitation exhibited would reflect the component that dominates the “race” to decision output (Logan, 1988; Marsolek, 2004). The division of abstract and specific representations as subserved by parallel left and right hemisphere pathways is consistent with this view (Marsolek, 2004). Alternatively, these components could be additive with the additional facilitation associated with multiple repetitions added to the abstract component associated with a single repetition. Future experiments will be required to resolve this issue.

The sensitivity of classification learning to the correspondence between study and test format indicates a high degree of specificity in the establishment and utilization of learned responses. Such hyperspecificity (Schacter, 1985) is consistent with the notion that response learning is episodic in nature. In fact, the working assumption is that instance learning reflects storage and retrieval of episodic representations (see Tenpenny, 1995 for a discussion of this issue). Similar to the results from Experiment 2, a number of studies have demonstrated that episodic representations encode visually specific information and thereby are sensitive to changes in visual format (Biederman & Cooper, 1991; Lawson, 2004; Srinivas & Verfaellie, 2000). In addition, our own work has already demonstrated that response learning is impaired in MTL amnesics (Schnyer, Dobbin, Nicholls, Schacter, & Verfaellie, 2006). Within the realm of repetition priming, however, the key question is whether response learning reflects conscious intentional retrieval of previous item-decision associations (Schacter, Bowers, & Booker, 1989) or an unconscious unintentional form of memory that nonetheless is MTL dependent and episodic in nature (Chun & Phelps, 1999). In this regard, evidence is beginning to emerge suggesting that response learning reflects an unintentional form of memory. In a series of studies exploring the laterality dissociations between abstract and specific visual representations in word stem completion priming, Marsolek (2004) examined both case-specific priming and case-specific explicit memory. He obtained a clear dissociation between the characteristics of the implicit and explicit tasks, indicating that visually specific priming cannot be attributed to explicit retrieval. Additionally, a recent electrophysiological study of response learning in our lab (Schnyer, Dobbins, Nicholls, Schacter, & Verfaellie, 2005) has demonstrated that prior response associations are retrieved at approximately 230 ms after stimulus presentation. Event-related potential research has revealed a reliable parietally located “old-new” effect that indexes recollection and occurs between 400 and 800 msec post stimulus onset (Gonsalves & Paller, 2000; and see Rugg, 1995 for a review). Therefore, the time course of response association retrieval appears considerably earlier than effects associated with conscious retrieval of episodic memories. Finally, intuition tells us that less effort would be required to perform the relatively simple object size judgment than would be involved in explicitly retrieving a prior classification (Jacoby, 1991) and then utilizing that information to guide responses. Final resolution of this question will require continuing experiments that seek clear explicit/implicit retrieval dissociations in response learning.

In addition to the decision specificity examined here, across the repetition priming literature there are a number of other instances of hyperspecificity in priming, including stimulus specificity and associative specificity (Schacter et al., 2004). Whether these other forms of specificity similarly reflect a form of decision learning remains to be determined, as does their dependence on the MTL (Marsolek, 2004). In this regard, it is useful to consider several characteristics of decision learning. First and most apparent is the inflexibility of the stimulus and decision components that become associated. Alteration of either of these components challenges the established association, thus disrupting learning. Second, in contrast to priming of abstract representations, decision learning is enhanced with multiple repetitions, although such learning may reach a maximal level of facilitation relatively quickly (Buckner, et. al, 1998; Logan, 1990). Finally, decision learning appears to depend on associative learning mechanisms mediated by the medial temporal lobes, as demonstrated by the fact that instance learning is impaired in patients with amnesia (Schnyer et al., 2006). Future studies can utilize these basic characteristics in order to determine the extent to which a similar learning mechanism is responsible for the specificity effects evident across a wide range of repetition priming paradigms.