工作記憶與分類作業之關連：針對特定分類規則及作業階段之檢驗

(1)

工作記憶與分類作業之關連：針對特定分類規則及作業階

段之檢驗

研究成果報告(精簡版)

計畫類別：個別型

計畫編號： NSC 99-2628-H-004-002-

執行期間： 99 年 08 月 01 日至 100 年 07 月 31 日

執行單位：國立政治大學心智、大腦與學習研究中心

計畫主持人：楊立行

計畫參與人員：碩士級-專任助理人員：鄭惟尹

碩士班研究生-兼任助理人員：吳岳勳

碩士班研究生-兼任助理人員：張俊彥

碩士班研究生-兼任助理人員：盧毓文

碩士班研究生-兼任助理人員：李孟潔

碩士班研究生-兼任助理人員：林家源

碩士班研究生-兼任助理人員：鍾德政

報告附件：出席國際會議研究心得報告及發表論文

公開資訊：本計畫可公開查詢

中華民國 100 年 11 月 01 日

(2)

及類別學習。自從 10 年前 COVIS 模型被提出後，研究者開始

積極探索工作記憶與類別學習之間的關係。尤其，由於 COVIS

模型主張唯有單向度的類別規則學習（RB），才需要工作記憶

的調控；雙向度的類別規則（II）則不需要。許多研究者都致

力於檢驗此一說法。然而，目前所得證據卻仍然相當分歧。一

派學者發現當工作記憶容量被擠壓時，RB 規則的學習表現會變

差；但，II 規則的學習則不受影響。另一派學者則是指出這類

的分離現象，很多都是由於作業難度沒控制好，或者其它混淆

變項所引起。這些分歧的主要原因之一有可能來自於它們都是

小樣本的實驗，以致於所觀察到的現象有可能含有極大的隨機

誤差成份。因此，本研究採用大量尺的研究方法，以超過 100

人的受試者樣本進行實驗。同時為了得到良好的難度控制，本

實驗共採用四種對稱控制的類別學習情境（RB-難，RB-易，

II-難，及 II-易）。此外，也針對每位受試者以四種工作記憶

廣度測驗進行工作記憶廣度的測量（memory updating,

operation span, sentence span, 及 spatial short-term

memory）。所得之數據再經由結構方程模式進行分析，檢驗是

否 RB 與 II 分屬不同的潛在變項，而且只有對應於 RB 的潛在

變項才與工作記憶廣度的潛在變項有相關。結果顯示，最佳的

結構方程模式應為只有一個單一潛在變項（類別學習潛在變

項）來代表全部的類別學習表現。即，RB 與 II 不需要分別用

不同的潛在變項表徵。在這個最佳的結構方程模式中，工作記

憶廣度的潛在變項與類別學習潛在變項之間相關高達.50。這

顯示不論是何種類別規則的學習都和工作記憶有相關，而工作

記憶廣度愈大者，學習表現也愈好。因此，本研究之結果並不

支持 COVIS 模型。

英文摘要： Working memory is crucial to many aspects of

cognition, including reading comprehension, reasoning,

and categorization. Since the emergence of COVIS model

\cite{Ashby_etal_1998}, the relationship between

working memory and category learning has attracted

researchers‘ attention. Specifically, as induced by

COVIS model, whether working memory exclusively

mediates the learning of uni-dimensional

categorization rule (i.e., RB) but not the learning of

multi-dimensional categorization rule (i.e., II)

becomes a major theoretical debate. Although the

advocators for COVIS model provide evidence for the

dissociation between the learning performance in RB

and II conditions and the increased load of working

(3)

possibility of the involvement of other confounding

variables. Among the possible reasons for these

inconsistent findings, the small sample size might be

relatively paramount. Therefore, in order to get a

better clarification, this study adopts a large-scale

method with N > 100, each of whom were tested under 4

fully counterbalanced conditions of two difficulties

and two rule types (i.e., RB and II). Also, the

participants‘ working memory capacity were measured

by four working memory capacity tests: memory updating

(MU), operation span (OS), sentence span (SS), and

spatial short-term memory (SSTM). The relationship

between working memory capacity and category learning

was then examined by structural equation modeling

(SEM). The SEM result consistently supports a strong

correlation between the latent variables for working

memory capacity and category learning (r=.5). Thus,

the present result shows that both the learning of RB

and II rules correlate with working memory capacity.

(4)

Running head: WORKING MEMORY AND CATEGORIZATION

NSC

99-2628-H-004-002-Working Memory Does Not Dissociate Between Different Perceptual Categorization Tasks

Lee-Xieng Yang

National Cheng-Chi University

Lee-Xieng Yang

Department of Psychology National Chengchi University

NO.64,Sec.2,ZhiNan Rd.,Wenshan District,Taipei City 11605, Taiwan (R.O.C) [email protected]

(5)

Abstract

Working memory is crucial to many aspects of cognition, including reading

comprehension, reasoning, and categorization. Since the emergence of COVIS model (Ashby, Alfonso-Reese, Turken, & Waldron, 1998), the relationship between working memory and category learning has attracted researchers’ attention. Specifically, as induced by COVIS model, whether working memory exclusively mediates the learning of uni-dimensional categorization rule (i.e., RB) but not the learning of multi-dimensional categorization rule (i.e., II) becomes a major theoretical debate. Although the advocators for COVIS model provide evidence for the dissociation between the learning performance in RB and II conditions and the increased load of working memory only influences the RB performance, not the II performance, other researchers point out the possibility of the involvement of other confounding variables. Among the possible reasons for these inconsistent findings, the small sample size might be relatively paramount. Therefore, in order to get a better clarification, this study adopts a large-scale method with N ¿ 100, each of whom were tested under 4 fully counterbalanced conditions of two difficulties and two rule types (i.e., RB and II). Also, the participants’ working memory capacity were measured by four working memory capacity tests: memory updating (MU), operation span (OS), sentence span (SS), and spatial short-term memory (SSTM). The relationship between working memory capacity and category learning was then examined by structural equation modeling (SEM). The SEM result consistently supports a strong correlation between the latent variables for working memory capacity and category learning (r=.5). Thus, the present result shows that both the learning of RB and II rules correlate with working memory capacity.

(6)

NSC

99-2628-H-004-002-Working Memory Does Not Dissociate Between Different Perceptual Categorization Tasks

Our ability to discriminate different objects and group them together in classes of similar entities—to categorize—is recognized as a fundamental aspect of human cognition (e.g., Estes, 1994). Likewise, the ability to hold and manipulate information in memory for brief periods of time—to use working memory—is also considered an elemental facet of human functioning (Oberauer, 2009).

Despite their status as two ‘pillars’ of cognition, until very recently, little research had focussed on the theoretical and empirical relationship between the two. This paper contributes to recent work that attempts to redress the balance (e.g., DeCaro, Thomas, & Beilock, 2008; Newell, Dunn, & Kalish, 2010).

There are several reasons for focussing on the relationship between working memory (WM) and category learning. First, exploring their empirical relationship may help adjudicate between competing theories of working memory—for example, whether it is best understood as executive attention (e.g., Kane, Bleckley, Conway, & Engle, 2001) or the ability to bind together temporary and transient representations (Oberauer, 2009). Second, examination of the relationship can shed light on the debate over single versus multiple-system views of category learning. Theoreticians who ascribe to the

multiple-systems view generally propose one “explicit” system that is reliant on WM (specifically for the storage and testing of verbal rules), and one “implicit” system that does not require WM but learns associations between (motor) responses and category labels (e.g., Ashby & Maddox, 2005). The product of learning from the latter system is often assumed to be unavailable to awareness or impossible to verbalize (Ashby et al., 1998).

(7)

Characterizing the two memory systems in this way reinforces the central

distinguishing role played by WM, and the importance of clarifying the exact nature of its contribution to category learning. Despite this centrality, much of the evidence for the involvement of WM comes from indirect measures such as examining the effects of concurrent cognitive tasks during learning (e.g., Waldron & Ashby, 2001). These studies typically compare the learning of tasks in which stimuli can be classified on the basis of simple verbalizable rules to those in which verbal rules cannot—so it is claimed—be readily applied (Minda & Miles, 2010). The latter tasks are commonly referred to as information-integration tasks.

To clarify our nomenclature, we will refer to the two types of tasks as RB

(rule-based) and II (information-integration), respectively, from here on. By contrast, we will be referring to people’s imputed cognitive strategies as “rule use” and “information integration,” respectively. That is, although people typically approach an RB task by applying a rule, it is also common for people to solve II tasks by resorting to the

application of one-dimensional rules; hence, it is sometimes advantageous to differentiate between the physical task parameters (RB vs. II) and the psychological process (rule use vs. information integration) that people engage to solve a task. This distinction will be clarified further during presentation of our data.

According to the multiple-systems view, only RB tasks, learned by the explicit system, require the resources of WM (Zeithamova & Maddox, 2006). II tasks, by contrast, when approached via information integration, are the domain of the implicit system and therefore are not reliant on WM. To date, several studies have reported results consistent with the selective involvement of WM in RB tasks using a dual task methodology (Foerde, Poldrack, & Knowlton, 2007; Waldron & Ashby, 2001; Zeithamova & Maddox, 2006). Briefly, those studies show that RB performance, but not II performance, is impaired by a secondary task that ostensibly occupies WM during training. However, in each of these

(8)

cases there are good reasons to question the “multiple-system” interpretation of the reported results (see, e.g., Nosofsky & Kruschke, 2002; Nosofsky, Stanton, & Zaki, 2005). Furthermore, as already noted, such studies can only provide indirect evidence about WM involvement because none of them attempted to actually measure the working-memory capacity (WMC) of participants.

Those few studies that have included direct measurement of both WMC and

category learning ability are similarly equivocal about the differential involvement of WM in RB and II tasks. For example, DeCaro et al. (2008) reported a dissociation between WMC and categorization performance, such that WMC was positively associated with performance on a rule-based task but negatively associated with performance on an information-integration task. However, this result has since undergone considerable re-evaluation: Tharp and Pickering (2009) suggested that the outcome may have reflected an inappropriate performance measure; namely an insufficient number of trials in a trials-to-criterion measure. Tharp and Pickering showed that this criterion was

unacceptably lax and thus might have spuriously created a negative correlation between WMC and information-integration performance. In a small-scale replication of their earlier study, DeCaro, Carlson, Thomas, and Beilock (2009) confirmed that WMC was positively associated with performance in both rule-based and information-integration tasks when a more suitable performance criterion was used.

In a related study, Minda, Desroches, and Church (2008) tackled the same

theoretical questions using a developmental approach: Children of various ages and adults were compared on their ability to learn different categorization problems, and adults were found to outperform children only on rule-based tasks but not on information-integration tasks. However, this result should be interpreted with some caution because children differ from adults in many ways other than development of their WM, and differences between age groups therefore only indirectly illuminate the role of WM.

(9)

Common to all studies addressing the role of WMC in categorization to date has been that stimuli comprised discrete binary dimensions that were either amenable to verbal rules or that were sufficiently discriminable to afford memorization as exemplars (Rouder & Ratcliff, 2004). In particular, even with multi-dimensional problems, partial rules can get you quite far and support above-chance performance (Tharp & Pickering, 2009). This attribute of the stimuli diffuses the demarcation between II and RB tasks because it facilitates a (partial) rule-based approach even to II tasks. For that reason, there has been a growing shift in the literature towards the use of more “perceptual” stimuli—such as Gabor Patches; see Figure 1 for sample stimuli—that have continuous dimensions and escape ready verbalization (e.g., Markman & Maddox, 2006; Zeithamova & Maddox, 2007). Gabor patches that must be classified on the basis of two dimensions arguably escape ready verbalization.

These perceptual stimuli have the additional advantage of permitting a systematic manipulation of task difficulty, which is the focus of the present study. Category

boundaries can be carefully constructed to partition regions of perceptual space and exemplars of each category can be placed at different distances from these boundaries. The distance-to-boundary permits fine-tuning of the ease with which exemplars from the different categories can be discriminated, permitting us not only to look at the relation between WMC and II and RB tasks, but also at the relation between WMC and ’easy’ and ’hard’ versions of the two tasks. The present article thus systematically explores the involvement of WMC in the two classes of tasks across several levels of difficulty.

The orthogonal manipulation of task type and difficulty in our design is one advance on the only study to date that has explored the link between WMC and performance with continuous-dimensional stimuli. Newell et al. (2010) found that participants with higher WMC performed more accurately on RB and II tasks. Although their results provide support for the general notion that WMC and perceptual categorization are positively

(10)

linked, Newell et al. (2010) had no within-experiment manipulation of task difficulty. This limitation precluded the more extensive analysis of the relationship between category learning and WMC that we offer here via structural equation modeling (SEM). SEM requires the presence of multiple indicator variables (e.g., multiple RB/II tasks performed by the same participants) in order to identify latent variables whose correlation with other constructs (e.g., WMC) can then be ascertained without contamination by measurement error and task idiosyncracies.

In sum, the experiment in this study provides a comprehensive assessment of the role played by WM in perceptual category-learning tasks that differ both in terms of structure (II vs. RB) and difficulty. In both experiments, learning performance was found to be related to WMC, and that relationship could be captured by a single latent variable in each study that represented performance on II and RB tasks together. The positive mediating role of WMC extended to an analysis in which each individual’s performance was expressed as the likelihood of having employed a rule-based or information-integration strategy: Although the preferred choice of strategy differed between tasks—with

one-dimensional rule use predominating in RB tasks and information-integration strategies being more prevalent in II tasks—increasing WMC facilitated use of both strategies. The data provide no evidence for the suggestion that WM is selectively involved in some types of perceptual category-learning but not others.

Experiment 1 Method Participants and Categorization Stimuli

One hundred and nineteen undergraduate students at National Chengchi University participated in this experiment. The participants were paid NT$100 (∼= US$3.50) for each session (about 1 hr). Every participant completed the two sessions on different days

(11)

within a two-week period. Additionally, in order to encourage use of the optimal

categorization boundary, a bonus of NT$50 (∼= US$1.75) was paid if accuracy in any one of the learning blocks exceeded a criterion (explained below). The bonus was not paid more than once for each category-learning task. Across the four tasks (2 types × 2 difficulties), the maximum achievable bonus was thus NT$400 (∼= US$14).

The stimuli in Experiment 1 were Gabor patches that varied along two dimensions, viz. orientation and spatial frequency. In this experiment, the 4 conditions (RB-MED, RB-HD, II-EZ, and II-MED) were extended by including an additional 44 novel transfer stimuli. The category structures are shown in Figure 2.

Procedure

WMC was again measured using the battery presented by Lewandowsky, Oberauer, Yang, and Eker (2010). All four WMC tasks were used here: An operation span task (OS), a sentence span task (SS), a memory updating task (MU), and a spatial short-term memory task (SSTM). Because all participants in this experiment were Chinese native speakers, the Chinese mode of the WMC battery was chosen for the SS task (for details, see Lewandowsky et al., 2010).

Each categorization task comprised two types of blocks: Training blocks of 66 trials each and transfer blocks of 44 trials involving stimuli not seen during any of the training blocks. Altogether there were 8 blocks for each task, with blocks 2, 5, and 8 being transfer blocks and the remainder training blocks. Training trials were administered in the same way as in Experiment 1. Transfer trials were identical except that no feedback was provided after a response.

Tasks were assigned to the two experimental sessions in two possible sequences. In Sequence 1, the tasks were spared across sessions as follows:

(12)

session 2. In Sequence 2, the assignment of tasks to sessions was:

II-EZ→SS→MU→RB-MED for session 1 and RB-HD→OS→SST→II-MED for session 2. Participants were randomly assigned to Sequence, with an equal number in each.

Results Working Memory Capacity

Summary statistics for the WMC tasks for all participants (N = 119) are shown in Table 1 and the correlations between them in Table 2.

Categorization

Average performance across blocks is shown in the left panel of Figure 3 for all 4 tasks. Visual inspection shows that the learning performance in the RB-MED, RB-HD, and II-EZ tasks overlapped considerably, as intended and as expected. The II-MED task led to worse performance (M = .69) than the others (M = .77). The final learning

performance of the II tasks was significantly better than what would be expected from the optimal bilinear alternative; for II-EZ (.85 vs. .76), t(118) = 8.57,p < .01, and for II-Med (.76 vs. .73), t(118) = 2.56, p < .05.

Performance on the transfer blocks was summarized by computing the proportion of “correct” responses across the three blocks, where each transfer response was considered correct when an item was classified as intended by the experimenter. The resultant means are shown in the right panel of Figure 3 together with the best performance that could be expected on the basis of application of a bilinear boundary.

It is obvious from the figure that people improved across repeated transfer blocks, which was confirmed in a 3 (test block) × 4 (task) within-subjects ANOVA,

F(2, 236) = 148.95, M Se = 15.64, p < .0001. Likewise, the obvious differences between tasks were significant, F (3, 354) = 12.95, M Se = 37.78, p < .0001, whereas the interaction

(13)

between the two variables was not, F (6, 708) < 1. The transfer data confirm the pattern evident in the training data, albeit with a much larger corpus of stimuli and in the

absence of feedback. We do not consider them further and instead focus on exploration of individual differences in the learning data.

Structural Equation Modeling

A measurement model was again constructed for WMC relying on the four indicator variables (OS, SS, MU, and SSTM). The loadings of the four tasks on the WMC latent variable are shown in Table ??. Following precedent (Ecker, Lewandowsky, Oberauer, & Chee, 2010; Lewandowsky et al., 2010), the correlation between the error terms for the span tasks, SS and OS, was allowed to be freely estimated (.23). Not surprisingly, the model fit extremely well, χ2_{(1) = .317, CFI = 1.0, RMSEA= 0.0 (90% CI 0.0 – 0.201),} SRMR= .0103.

For the categorization tasks, each task was again represented by two manifest variables with freely-estimated correlations between the error terms that represented odd and even trials, respectively. We compared two models: The first one involved a separate factor for RB and II; this model fit very well, χ2_{(15) = 14.412, CFI = 1.0, RMSEA= .0} (90% CI .0 – .084), SRMR= .0426. The second model used a single factor for all

categorization tasks. The fit of this model differed neglibibly, ∆χ2_{(1) = .323, from that of} the two-factor model, χ2_{(16) = 14.735, CFI = 1.0, RMSEA= .0 (90% CI .0 – .079),}

SRMR= .0433. We therefore combined the more parsimonious single-factor model with the WMC measurement model into a final structural model, which is shown in Figure 4. In replication of Experiment 1, WMC was again strongly (r = .50, p < .0001) and uniformly related to learning performance in all tasks.

(14)

Discussion

Experiment 1 replicated the finding in my previous NSC granted project almost exactly: Individual variation in performance on all four tasks was captured by a single latent variable which was associated with WMC. People’s WMC also predicted the extent to which they settled on a strategy with which to classify the items. There were, however, a few subtle differences from the outcome of the previous finding: First, unlike in the previous study, evidence for the trade-off between rule use and information integration on the II tasks was more diffuse. Second, for the RB-MED task, there was evidence that people relied equally on rule and information integration, as revealed by the fact that (a) a model in which all task-relevant loadings were constrained to be equal fit acceptably only when those two error terms were correlated, and (b) the loading for information integration (RB-MED-II indicator variable) was significantly greater than zero. Neither of those discrepancies calls into question the main results, which is that WMC determines performance in all tasks and also determines the extent to which people adopt an appropriate strategy.

General Discussion

At first glance, our data seem at odds with results that support a distinction between RB and II tasks, and in particular existing data that point to an adverse role of WMC in learning of II tasks (DeCaro et al., 2008). However, those results have already undergone considerable re-evaluation (e.g., DeCaro et al., 2009; Newell et al., 2010; Tharp & Pickering, 2009), and our data provide a fairly strong capstone to this re-evaluation. The sum total of available evidence now points in the opposite direction; namely, that working memory is uniformly and positively associated with category learning

(15)

Note that our results and the experiments just reviewed do not speak to the multiple other variables that have been said to dissociate RB and II performance (for a review, see Ashby & Maddox, 2005). The conclusions therefore do not pertain to the multiple-systems view overall but only to the role of working memory vis-`a-vis those two systems—nonetheless, it must be noted that a more comprehensive critique of the multiple-systems view of category learning has recently been put forward.

Turning to the role of WM in category learning, we first note that there is much evidence in other domains that WM is closely related to retention and learning over the long term. For example, Mogle, Lovett, Stawski, and Sliwinski (2008) recently argued that the long-term memory involvement in WM tasks was responsible for the known strong relationship between WMC and fluid intelligence. Similarly, Unsworth and Engle (2007) showed that WMC predicts performance in measures of recollection that are typically taken to reflect long-term memory involvement. Indeed, the claim that WMC might have an adverse effect on long-term performance (DeCaro et al., 2008) is surprising and at odds with much previous research, whereas the finding of a uniformly facilitative effect is consonant with previous research.

How, then, does WMC support category learning? Two alternative mechanisms can be cited: On the one hand, WMC could facilitate memory for specific exemplars, such that people with high WMC are able to form more lasting or more exact memories of instances. On the other hand, WMC might facilitate faster learning of some other task-relevant representation (e.g., a category boundary). Several lines of evidence cast doubt on the first possibility. For example, in a thorough comparison of exemplar models and rule-based theories of perceptual categorization, Rouder and Ratcliff (2004) showed that for stimuli and tasks similar to those used in our experiment, people are unlikely to rely on exemplar memory. Exemplar models tend to characterize performance only when there are few and highly discriminable stimuli. Moreover, even when performance can be

(16)

characterized by exemplar models, as in the study of the Shepard tasks by Lew (2011), individual differences turn out not to be related to the precision of exemplar memory. Lew (2011) showed that the “specificity” parameter within an exemplar theory (ALCOVE; Kruschke, 1992) was unrelated to WMC.

We therefore favor the second option, namely that WMC facilitates speed of learning of whatever representation underlies categorization in this instance. In support, Lew (2011) showed that WMC was related to the speed with which weights were updated in two neural-network models (one of which did not involve memory for exemplars). We therefore suggest that WMC expedites the formation of long-term associations—a notion that is entirely consistent with theories of WM that emphasize associative “binding” processes (e.g., Oberauer, 2009).

In summary, we offer a rather unequivocal empirical contribution: WMC is strongly related to all manifestations of perceptual category learning. We find no evidence for a dissociation involving the putative distinction between “information-integration” and “rule-based” tasks and their presumed underlying memory systems. Our data form a distinct benchmark for further theorizing that relates categorization and working memory, two acknowledged pillars of human cognition.

(17)

References

Ashby, F. G., Alfonso-Reese, L. A., Turken, A. U., & Waldron, E. (1998). A

neuropsychological theory of multiple systems in category learning. Psychological Review, 105 , 442-481.

Ashby, F. G., & Maddox, W. T. (2005). Human category learning. Annual Review of Psychology, 56 , 149-178.

DeCaro, M. S., Carlson, K. D., Thomas, R. D., & Beilock, S. L. (2009). When and how less is more: Reply to tharp and pickering. Cognition, 111 , 415-421.

DeCaro, M. S., Thomas, R. D., & Beilock, S. L. (2008). Individual differences in category learning: Sometimes less working memory capacity is better than more. Cognition, 107, 284-294.

Ecker, U. K. H., Lewandowsky, S., Oberauer, K., & Chee, A. E. H. (2010). The components of working memory udpating: An experimental decomposition and individual differences. Journal of Experimental Psychological: Learning, Memory and Cognition, 36 , 170-189.

Estes, W. K. (1994). Classification and categorization. New York: Oxford University Press.

Foerde, K., Poldrack, R. A., & Knowlton, B. T. (2007). Secondary-task effects on classification learning. Memory and Cognition, 35 , 864-874.

Kane, M. J., Bleckley, M. K., Conway, A. R. A., & Engle, R. W. (2001). A

controlled-attention view of working-memory capacity. Journal of Experimental Psychological: General, 130 , 169-183.

Kruschke, J. K. (1992). Alcove: An exemplar-based connectionist model of category learning. Psychological Review , 99 , 22-44.

Lew. (2011). Working memory capacity and categorization: Individual differences and modeling. Journal of Experimental Psychological: Learnin, 37 , 720-738.

(18)

Lewandowsky, S., Oberauer, K., Yang, L.-X., & Eker, U. K. (2010). A working meory test battery for matlab. Behavioral Research Method , 42 , 571-585.

Markman, A. B., & Maddox, W. T. (2006). Choking and excelling under pressure. Psychological Science, 17 , 944-948.

Minda, J. P., Desroches, A. S., & Church, B. A. (2008). Learning rule-described and non-rule-described categories: A comparison of children and adults. Journal of Experimental Psychological: Learning, Memory and Cognition, 34 , 1518-1533. Minda, J. P., & Miles, S. J. (2010). The inference of verbal and nonverbal processing on

category learning. In B. H. Ross (Ed.), The psychology of learning and motivation: Advances in research and theory (vol 52) (p. 117-162). San Diego: Academic Press. Mogle, J. A., Lovett, B. J., Stawski, R. S., & Sliwinski, M. J. (2008). Whats so special

about working memory? Psychological Science, 19 , 1071-1077.

Newell, B. R., Dunn, J. C., & Kalish, M. L. (2010). The dimensionality of perceptual category learning: A state-trace analysis. Memory and Cognition, 38 , 563-581. Nosofsky, R. M., & Kruschke, J. K. (2002). Single-system models and inference in

category learing: Commentary on waldron and ashby (2001). Psychonomic Bulletin & Review, 9 , 169-174.

Nosofsky, R. M., Stanton, R. D., & Zaki, S. R. (2005). Procedural interference in perceptual classification: Implicit learning or cognitive complexity? Memory and Cognition, 33 , 1256-1271.

Oberauer, K. (2009). Design for a working memory. In B. H. Ross (Ed.), (p. 45-100). Rouder, J. N., & Ratcliff, R. (2004). Comparing categorization models. Journal of

Experimental Psychological: General, 133 , 63-82.

Tharp, I. J., & Pickering, A. D. (2009). A note on decaro, thomas, and beilock (2008): Further data demonstrate complexityies in the assessment of information-integration category learning. Cognition, 111 , 410-414.

(19)

Unsworth, N., & Engle, R. W. (2007). The nature of individual differences in working memory capacity: Active maintance in primary memory and controlled search from secondary memory. Psychological Review , 114 , 104-132.

Waldron, E. M., & Ashby, F. G. (2001). The effects of concurrent task interference on category learning: Evidence for multiple category learning systems. Psychonomic Bulletin & Review, 8 , 168-176.

Zeithamova, D., & Maddox, W. T. (2006). Dual-task interference in perceptual category learing. Memory and Cognition, 34 , 387-398.

Zeithamova, D., & Maddox, W. T. (2007). The role of visuospatial and verbal working memory in perceptural category larning. Memory and Cognition, 35 , 1380-1398.

(20)

Author Note

Preparation of this paper was facilitated by a Grant from National Science Council (NSC99-2628-H-004-002).

(21)

Table 1

Summary of Working Memory Capacity (WMC) scores in Experiment 1

Measure MU OS SS SSTM M .82 .79 .83 .89 s (.14) (.12) (.09) .(14) Minimum .30 .16 .44 .77 Maximum 1.00 .96 1.00 1.00 Skewness −1.15 −1.63 −1.00 0.05 Kurtosis 1.22 5.05 0.93 −0.13 Cronbach’s α .80 .80 .80 .93 Standardized loadings .64 .52 .57 .47

Note: MU = Memory updating task; OS = Operation span task; SS = Sentence span task; SSTM = spatial short-term memory. M = mean; s = standard deviation; Standardized loadings refer to WMC measurement model.

(22)

Table 2

Correlations between WMC tasks in Experiment 1

MU OS SS MU OS .32** SS .37** .46** SSTM .30** .27** .25** ∗ = p < .05, ∗∗ = p < .01

(23)

Figure Captions

Figure 1. Four sample Gabor patches used as stimuli in the present experiments.

Figure 2. The category structures for all 4 tasks used in Experiment 1, with

information-integration tasks (II) in the top row of panels and rule-based tasks (RB) in the bottom row. Open circles and squares refer, respectively, to the training stimuli of the two categories. The crosses represent transfer stimuli shown during the transfer blocks only and without response feedback.

Figure 3. Left panel: Proportion correct across training blocks for all categorization tasks in Experiment 2. Blocks are numbered by their ordinal position in the sequence, with transfer blocks (in positions 2, 5, and 8) omitted. Filled plotting symbols are for information-integration (II) tasks, and open plotting symbols are for rule-based (RB) tasks. Right panel : Proportion correct across transfer blocks for all categorization tasks in Experiment 2. Blocks are numbered by their ordinal position in the sequence, with

training blocks omitted. Filled plotting symbols are for information-integration (II) tasks, and open plotting symbols are for rule-based (RB) tasks. The data points for “Opt Alt” refer to the best performance that could be expected on the basis of application of two-dimensional bilinear boundaries.

Figure 4. Structural model relating working memory capacity (latent variable WMC ) to category learning performance in Experiment 1. All loadings and correlations are

standardized estimates. Manifest variables for category learning represent proportions correct on odd (terminal “–O”) and even (“–E”) trials for rule-based (“RB–”) and information integration (“II–”) tasks. The difficulty of tasks is coded as easy (“–EZ–”), medium (“–MED–”) or hard (“–HD–”). All loadings and coefficients shown are

(24)

(25)

W or k in g M em or y an d C at egor iz at ion , F igu re 7 0 1 2 3 4 5 6 0 45 90 Oreientation II−Easy 0 1 2 3 4 5 6 0 45 90 II−Medium 0 1 2 3 4 5 6 0 45 90 Frequency RB−Hard 0 1 2 3 4 5 6 0 45 90 Frequency Orientation RB−Medium

(26)

W or k in g M em or y an d C at egor iz at ion , F igu re 8 0.0 0.2 0.4 0.6 0.8 1.0 Learning Performance Block Pr(Correct) 1 3 4 6 7 RB−Med RB−HD II−EZ II−Med 0.0 0.2 0.4 0.6 0.8 1.0 Transfer Performance Block Pr('Correct') 2 5 8 Opt Alt RB-Med RB-HD II-EZ II-Med

(27)

Working Memory and Categorization, Figure 9 Learning .92 .91 .45 .46 .63 .64 .54 .63 .70 .91 .84 .85 .61 .54 .50 .50 .52 .25

(28)

本次赴澳洲墨爾本大學參加 2011 年澳洲數學心理學年會（Australian

Mathematical Psychology），主要是發表本人與碩士生之研究成果。該研究針

對兩個著名的類別分習模型（ALCOVE 和 SUSTAIN）進行比較。比較兩模型時，

有別於以往僅以單筆資料檢核模型對資料的適配性，本研究針對參數空間內各

種參數的組合，進行大範圍的檢查，針對每種參數組合，兩模型均做出預測，

然後再計算所做出的預測有多少比例會出現違反該模型自身假設，並以此比例

說明兩模型中孰優孰劣。研究結果顯示 ALCOVE 模型的內在邏輯一致性要強於

SUSTAIN 模型，即，SUSTAIN 模型可能會因為太過彈性，而對資料造成過度適配

之情形（over fitting）。這種比較理論的方式在實驗心理學中並不多見，於

會議報告時，與會人士多半給予正面肯定。除此之外，也有與會人士如

Stephan Lewandowsky 以及 Daniel Little 指出，或許這不能算是 SUSTAIN 的

缺失，而只能算是該模型之特徵之一，並經討論建議本人再次檢查 ALCOVE 模型，

以確保上述比較的穩定性。本次會議的其它收穫包括：邀請美國路易斯安那大

學的教授 Mike Kalish 來台進行 Bayesian inference 的工作坊，預計明年會實

施，以及，獲得新興期刊 psychology 的邀稿，令人欣慰。

(29)

Fwd: AMPC 2011 Acceptance

楊立行 <[email protected]> 28 January 2011 14:42

To: [email protected]

Forwarded message

---From: David Keisuke Sewell <[email protected]> Date: 2011/1/25

Subject: AMPC 2011 Acceptance To: 楊立行 <[email protected]>

Dear Lee-Xieng,

I am writing to confirm that your talk proposal has been accepted for the 2011 Australiasian Mathematical Psychology Conference. We look forward to seeing you in Melbourne in February! I will be in touch later this week with a draft schedule for the meeting.

All the best,

Dave

David Sewell

Postdoctoral Research Fellow Psychological Sciences University of Melbourne Victoria 3010

Tel: +61 (03) 8344 8156 Fax: +61 (03) 9347 6618

(30)

Similarity change by attention: Model

comparison between ALCOVE and SUSTAIN

Lee-Xieng Yang1 _{and Chung-Yu Wang}2

1_{National Chengchi University}

2_{National Cheng-Kung University}

Contact: [email protected]

Abstract

Most exemplar models for categorization postulate that a negative ex-ponential function transforms distance between objects in psychological space into similarity. There are two different ways to make this trans-formation. The first (Type I) is adopted by most exemplar models, such as GCM and ALCOVE, whereby distance on all stimulus dimensions is combined before being transformed to similarity. The second (Type II) is adopted by SUSTAIN, whereby distance on each stimulus dimension is transformed to dimensional similarity, before being combined across di-mensions to yield whole-object similarity. These two ways, however, might lead to very different similarity computations, specifically when one ob-ject is nearer to the target than another on one dimension but farther on the other dimension, with the constraint that one object has a shorter spatial distance than the other to the target. This study with simula-tions of ALCOVE and SUSTAIN compares these two computasimula-tions by observing the rank ordered similarity of two objects to a common target. The results show that changes in dimensional attention do not lead to changes in rank ordered similarity comparisons in ALCOVE, which uses the Type I computation. However, rank ordered similarity does change in SUSTAIN, which uses the Type II computation. This is because in SUSTAIN, decay on dimensional similarity is weighted by dimensional attention such that similarity along the dimension nearest to the target decays more quickly than the similarity along the dimension farthest from the target. Although with this characteristic, SUSTAIN can predict the rule-plus-exemplar result in Erickson and Kruschke (1998), this implies that SUSTAIN is too sensitive to dimensional similarity.

(31)

日期:2011/11/01

國科會補助計畫

計畫名稱: 工作記憶與分類作業之關連：針對特定分類規則及作業階段之檢驗計畫主持人: 楊立行計畫編號: 99-2628-H-004-002- 學門領域: 實驗及認知心理學

無研發成果推廣資料

(32)

計畫主持人：

楊立行

計畫編號：

99-2628-H-004-002-計畫名稱：