ZEITHAMOVA DUAL TASK IN CATEGORY LEARNING 1 Dual Task Interference in Perceptual Category Learning

The effect of a working memory demanding dual task on perceptual category learning was investigated. In Experiment 1, participants learned unidimensional rule-based or information-integration category structures. In Experiment 2, participants learned a conjunctive rule-based category structure. In Experiment 1, unidimensional rule-based category learning was disrupted more by the dual working memory task than was information-integration category learning. In addition, rule-based category learning differed qualitatively from information-integration category learning in yielding a bimodal rather than a normal distribution of scores. Experiment 2 showed that rule-based learning can be disrupted by a dual working memory task even when both dimensions are relevant for optimal categorization. The results support the notion of at least two systems of category learning: a hypothesis-testing system that seeks verbalizable rules and relies on working memory and selective attention, and an implicit system that is procedural–learning-based and is essentially automatic. Introduction Humans live in a world of categories rather than unique instances. Categories divide the world into meaningful pieces. Humans categorize to reach cognitive economy of memory, to communicate and understand, to explain and predict properties and actions of new stimuli on the basis of older experiences. Because categorization is essential for higher level cognition, much attention in cognitive research has been paid to category learning (e.g. Ashby & Maddox, 2005; Kruschke, 1992; Estes, 1994; Love, Medin, & Gureckis, 2004; Medin & Schaffer, 1978; Nosofsky, 1986). A large and growing body of research suggests that participants have available multiple processing modes that can be used during categorization. Well established in the literature is a distinction between categorization according to a rule versus categorization based on the overall similarity (Allen & Brooks, 1991; Erickson & Kruschke, 1998; Folstein & Van Petten, 2004; Kemler Nelson, 1984; Nosofsky, Palmeri, & McKinley, 1994; Rehegr & Brooks, 1993). Building upon this work on multiple processing modes is a recent interest in understanding the neurobiological underpinnings of category learning and examining the possibility of multiple systems of category learning (Ashby, Alfonso-Reese, Turken, & Waldron, 1998; 1 This research was supported in part by National Institute of Health Grant R01 MH59196 to WTM. We thank Evan Heit, Sarah Queller and two anonymous reviewers for useful comments on an earlier version of this manuscript. We are especially indebted to Larry Cormack for his guidance on the statistical analyses. Correspondence concerning this article should be addressed to Dagmar Zeithamova or W. Todd Maddox, University of Texas, 1 University Station A8000, Department of Psychology, Austin, Texas, 78712 (e-mail: [email protected] or [email protected]). ZEITHAMOVA DUAL TASK IN CATEGORY LEARNING 2 Poldrack, Prabhakaran & Seger, 1999, Reber, Stark, & Squire, 1998; E. E. Smith, Patalano, & Jonides, 1998; for reviews, see Keri, 2003; and Maddox & Ashby, 2004). Relevant to this work are studies of multiple memory systems (Poldrack & Packard, 2003; Schacter & Tulving, 1994; Squire, 1992) and multiple reasoning systems (Sloman, 1996). One multiple systems model of perceptual category learning, and the only one that specifies the underlying neurobiology, is the COmpetition between Verbal and Implicit Systems (COVIS) model proposed by Ashby et al. (1998; Ashby & Waldron, 2000). COVIS postulates two systems that compete throughout learning – an explicit, hypothesis-testing system that uses logical reasoning and depends on working memory and executive attention, and an implicit, procedural learning-based system. [Relations between COVIS and the multiple process (rule vs. overall similarity) approach are reserved for the General Discussion.] At the implementation level, the explicit, hypothesis-testing and the implicit, procedural-learning systems have distinct, but partially overlapping neurobiological underpinnings. The key neural structures for the hypothesis-testing system are the prefrontal cortex, anterior cingulate, and the head of the caudate nucleus. The key neural structures for the procedural learning system are the inferotemporal cortex and the tail of the caudate nucleus. A dopamine-mediated reward signal from the substantia nigra is critical for learning in this system. Both systems attempt to acquire and solve every categorization task encountered. However, the relative weight of each system in the category judgment depends on the relative success of each system in category learning which in turn depends on the type of category structure to be acquired. The hypothesis-testing system searches for and applies explicit rules that are typically easy to verbalize (the hypothesis testing system is often called verbal, although such a description may not be appropriate in all cases and is not appropriate for nonhumans). One example is a unidimensional rule that allows categorization on the basis of a criterion along a single separable stimulus dimension. For example, if the stimulus is a Gabor patch (a sinusoidal luminance grating windowed by a Gaussian envelope) that varies across trials in spatial frequency and spatial orientation, one unidimensional rule may be: “respond A if the spatial frequency is low, and respond B if the spatial frequency is high.” Category structures for which the optimal rule is likely to be explicitly verbalized as a categorization strategy by a responder are called rule-based. The implicit, procedural learning-based system learns to associate a category response with a region of perceptual space without deriving any explicit rule. Stimuli are represented perceptually in higher order visual areas such as inferotemporal (IT) cortex. It is well established that there is a many-to-one convergence of IT cells onto each cell in the tail of the caudate (Wilson, 1995). The striatum learns to associate subregions in the perceptual space with category assignments (Packard & Knowlton, 2002). Category structures acquired by the implicit system may be very complex (e.g. Ashby & Maddox, 1992, 2005). The categorization rule typically combines two or more stimulus dimensions expressed in different units. Such a rule may be: “respond A if the spatial frequency is greater than the spatial orientation, otherwise respond B” A person that categorizes according such a rule is not likely to be able to verbalize it in this form because it compares values expressed in different units and thus is not easy to comprehend logically. We say that the optimal rule is not verbalizable. Category structures in which the optimal rule is of this form are called information-integration. As a consequence of the proposed underlying neurobiology, the nature of the feedback and response mapping should affect the implicit, procedural learning-based system, but not the hypothesis testing system, and thus should affect information-integration but not rule-based category learning. Indeed, several studies have reported results consistent with this prediction (Ashby, Queller & Berretty, 1999; Maddox, Ashby, & Bohil, 2003; Maddox & Ing, 2005; Maddox, Bohil & Ing, 2004). On the other hand, working memory load should affect the explicit, hypothesis testing system, but not the implicit, procedural 2 Using the term “implicit” for the procedural learning based system does not imply an unconscious nature to category learning by this system (see Shanks and St. John, 1994 for a discussion of unconscious learning). Rather, we mean that the optimal categorization rule characterizing the information-integration category structure is not directly represented in the system, but only implicitly present in the stimulus-response mapping within striatum. The participants are not likely to be able to express the rule explicitly after the training, even when their response strategies suggest that they are able to employ such a rule. ZEITHAMOVA DUAL TASK IN CATEGORY LEARNING 3 learning-based system, and thus should affect rule-based but not information-integration category learning. Waldron and Ashby (2001) found support for this prediction using a working memory demanding dual task. Specifically, they found a large dual task interference on uni-dimensional rule-based category learning, but only a small dual task interference on (multidimensional) information-integration category learning when a small number of highly discriminable binary-value dimension stimuli were used. The goal of the current research was two-fold. First, and foremost, we wished to test the generality of Waldron and Ashby’s (2001) results when applied to a uni-dimensional rule-based and (twodimensional) information-integration category learning task using a large number of perceptually similar continuous-value dimension stimuli—that is, Gabor patches that varied across trials in spatial frequency and spatial orientation. Stimuli of this sort have been used extensively to study category learning (see Maddox & Ashby, 2004 for a review). Second, we explored the dual task interference phenomenon in more detail, and provided a critical test of Nosofsky and Kruschke’s (2002; see also Ashby & Ell, 2002) single system explanation of the original Waldron and Ashby results, by examining two-dimensional, conjunctive rulebased category learning in a dual task setting. In the next section, we briefly review a number of empirical studies that test a priori predictions from COVIS with a more detailed review of Waldron and Ashby (2001). Then we present the results from two Experiments. We conclude with some general comments that include a discussion of alternative approaches to categorization and how they may or may not account for the experimental results. Brief review of COVIS and the dissociation studies COVIS assumes that, regardless of the nature of the category structures (i.e., rule-based or information-integration), both the hypothesis testing system and the procedural-learning system attempt to learn. The two systems then compete to determine the response. COVIS assumes an initial bias for the explicit system. If an explicit rule exists that yields good performance, the hypothesis-testing system is likely to be successful and dominate the implicit system. If no such explicit rule exists, the hypothesistesting system will continuously fail to discover the correct rule and the implicit system will eventually dominate. To study the properties of each system, different category structures are therefore used: rulebased category structure for studying the explicit system and information-integration category structure for studying the implicit system. Evidence for multiple processes involved in category learning comes from a number of sources (see Maddox and Ashby, 2004 for a review). Several experiments have shown that information-integration, but not rule-based category learning, may be disrupted by feedback or instruction manipulation. First, category learning is qualitatively different under trial-by-trial feedback than it is without feedback. Without supervision, people typically use simple unidimensional rules (Ashby, Queller & Berretty, 1999), whereas with trial-by-trial feedback they are able to learn complex non-linear decision bounds that are difficult to describe verbally (Ashby & Maddox, 1992). Second, when the feedback is delayed, processing in the implicit system is affected, so that learning of an information-integration category structure may be impossible (Maddox, Ashby, & Bohil, 2003; Maddox & Ing, 2005). Third, most experiments use consistent stimulus-response mappings. For example, when a stimulus is presented, the participant is asked to press button “A” with the left hand and button “B” with the right hand (A-B training). Maddox, Bohil and Ing (2004; see also Ashby, Ell, & Waldron, 2003) used a variable stimulus-response mapping. The participants were asked to press either button “Yes” or button “No” to a stimulus in response to a question “Is this an A?” or “Is this a B?” (Yes-No training). As predicted by COVIS, Yes-No training impaired informationintegration category learning compared to A-B training, but had no effect on the rule-based category learning. The previous studies show that the implicit system differs from the hypothesis-testing system in that it requires immediate feedback and a consistent stimulus-response mapping. When feedback or a consistent stimulus-response mapping is not provided, learning by the implicit system is adversely affected. COVIS postulates that this is due to the fact that learning in the implicit system is dopamine-mediated. Positive feedback induces dopamine to be released from the substantia nigra into the tail of the caudate nucleus, strengthening recently activated synapses. When the feedback is not provided or is substantially delayed synaptic activation within striatum decays and learning does not occur (Arbuthnott, Ingham, & Wickens, 2000; Kerr, & Wickens, 2001). ZEITHAMOVA DUAL TASK IN CATEGORY LEARNING 4 One may argue that the disruption of information-integration category learning but not rule-based learning is due to differences in complexity and therefore difficulty of simple (e.g. one-dimensional) verbalizable rules in the rule-based condition versus complex (multi-dimensional) nonverbalizable integration rules in the information-integration condition. To provide evidence for the existence of two alternative systems, double dissociation should be demonstrated. Recent studies introduced manipulations that impair rule-based category learning but not information-integration category learning (Maddox, Filoteo, Hejl & Ing, 2004; Maddox, Ashby, Ing & Pickering, 2004). Waldron & Ashby (2001) provided empirical evidence of that kind by introducing a second task to be performed concurrently with category learning. Review of Waldron and Ashby (2001) Recall that COVIS postulates that the hypothesis-testing system relies on working memory and selective attention to solve rule-based category tasks, whereas learning in the procedural-learning system is essentially automatic. Waldron and Ashby (2001) provided an empirical test of this prediction by comparing rule-based and information-integration category learning under dual task conditions with that in a single-task control. They chose a numerical analog of the Stroop task (for a detailed review of the Stroop task see MacLeod, 1991) to serve as a dual task. The Stroop task is known to require working memory and selective attention, and to strongly activate the anterior cingulate and prefrontal cortex (Bench at al., 1993), neural structures associated with the explicit, hypothesis-testing system, but not with the implicit procedural-learning system proposed in COVIS. Waldron and Ashby had participants learn to categorize colored geometric figures presented on colored background that varied on four binary dimensions. In the uni-dimensional rule-based condition, one dimension was relevant and the remaining three were irrelevant. In the information-integration condition, information from three dimensions had to be integrated and one dimension could be ignored (see Waldron & Ashby, 2001 for details). Under control conditions, the participant simply categorized each stimulus on every trial. In the dual task conditions, the participant had to perform a numerical analog of the Stroop task during each trial of categorization. The Stroop task stimulus was presented simultaneously with the categorization stimulus for 200 ms. The Stroop stimulus was then masked and the categorization stimulus remained on the screen until the participant categorized it. After categorization feedback, the participant was to respond to the Stroop stimulus they had seen at the beginning of the trial. Therefore, the participant was required to hold a representation of the Stroop stimulus in working memory during the process of categorization. Performance in the Stroop task was emphasized over the categorization task. Waldron & Ashby (2001) found that the dual task produced greater interference for the unidimensional rule-based task than for the information-integration task. These findings support the COVIS prediction that a dual working memory task impaired rule-based, but not information-integration category learning, and argues against the “complexity” arguments offered against multiple systems theories. EXPERIMENT 1 The main aim of Experiment 1 was to test the generalizability of Waldron and Ashby’s (2001) results in an experiment using a large number of unique continuous-valued dimension stimuli. The stimuli were Gabor patches that varied across trials in spatial frequency and spatial orientation. Unidimensional (UD) rule-based and information-integration (II) category learning were examined under control and dual Stroop conditions. Scatterplots of the stimuli used in the UD and II category learning conditions are shown in Figure 1 along with the optimal decision bound. Each point in the scatterplot denotes the spatial frequency and spatial orientation of a single stimulus. In the unidimensional condition, spatial frequency was relevant and spatial orientation was irrelevant and the optimal rule required participants to respond A when the spatial frequency was low and to respond B when the spatial frequency was high. Both dimensions were relevant in the information-integration condition. The optimal rule required participants to respond A when the difference of the value on spatial frequency dimension and the value on the spatial orientation dimension was low and to respond B when the difference of the values on the two dimensions was high. Such a rule is not easy to comprehend logically because it compares values in different units. The 3 Some of the studies reported above utilized multidimensional rule-based tasks that address this potential shortcoming and continued to show the predicted results (e.g., Maddox & Ing, 2005; Maddox, Bohil, & Ing, 2004). ZEITHAMOVA DUAL TASK IN CATEGORY LEARNING 5 category discriminabilities (d’) were 4.5 for unidimensional and 10.3 for information-integration category structure.