راهبردهای یادگیری تبعیض متمایز و ارتباط آنها با حافظه فضایی و کنترل توجه در 4- تا 14 ساله

Abstract Behavioral, psychophysiological, and neuropsychological studies have revealed large developmental differences in various learning paradigms where learning from positive and negative feedback is essential. The differences are possibly due to the use of distinct strategies that may be related to spatial working memory and attentional control. In this study, strategies in performing a discrimination learning task were distinguished in a cross-sectional sample of 302 children from 4 to 14 years of age. The trial-by-trial accuracy data were analyzed with mathematical learning models. The best-fitting model revealed three learning strategies: hypothesis testing, slow abrupt learning, and nonlearning. The proportion of hypothesis-testing children increased with age. Nonlearners were present only in the youngest age group. Feature preferences for the irrelevant dimension had a detrimental effect on performance in the youngest age group. The executive functions spatial working memory and attentional control significantly predicted posterior learning strategy probabilities after controlling for age.

Introduction The ability to learn from feedback is crucial in a changing environment. Using various paradigms in which learning from positive and negative feedback is essential, behavioral, psychophysiological, and neuropsychological studies have revealed large developmental differences in performance. Tasks used in these paradigms include a rule search and application task (van Duijvenvoorde et al., 2008 and Zanolie et al., 2008), a rule switch task (Crone, Zanolie, van Leijenhorst, Westenberg, & Rombouts, 2008), the Wisconsin Card Sorting Test (WCST) (Heaton, Chelune, Talley, Kay, & Curtis, 1993), and the discrimination learning task (e.g., Block et al., 1973, Kendler, 1979 and Raijmakers et al., 2001). These tasks have in common that one or more underlying rules need to be inferred from feedback and that the correct solution may be found by testing hypotheses. The results of the studies mentioned above suggest the presence of distinct modes of learning and feedback processing. In the current study, we used mathematical learning models to distinguish different learning modes on a discrimination learning task and to identify underlying strategies. Our specific aim was to examine the relation between these modes and the executive functions working memory and attentional control. In addition, we investigated the effect of preferences for stimulus features on learning performance. All of the experimental rule learning tasks mentioned above can be solved by applying hypothesis testing strategies. The tasks differ, inter alia, in the size of the set of possible rules, the number of rule shifts, the presence of ambiguous feedback, and whether the set of possible rules is known to the participants. For instance, on the WCST, a series of unidimensional card sorting rules need to be inferred (from feedback) and applied. Children typically perform worse than adults on a number of measures on the WCST, including the number of perseverative errors (Chelune and Baer, 1986, Heaton et al., 1993 and Huizinga et al., 2006). In a discrimination learning task, a simple, unidimensional categorization rule needs to be learned from positive and negative feedback (e.g., Kendler, 1979). The set of rules is not explicitly mentioned. In developmental studies using discrimination learning tasks, two distinct modes of learning have been observed—a fast and a slow learning mode—with an age-related increase in the probability of using the fast mode (e.g., Kendler, 1979 and Raijmakers et al., 2001). In her levels-of-functioning theory, Kendler (1979) posited that learning in the slow mode is incremental (based on associative stimulus–response learning), whereas the fast mode is based on a hypothesis testing strategy. However, the support for this theory is ambiguous (see Esposito, 1975, for a review). There is some evidence supporting the interpretation of the fast mode as a strategy of efficient hypothesis testing (e.g., Block et al., 1973, Kendler, 1979 and Raijmakers et al., 2001). The interpretation of the slow mode in terms of a well-defined strategy is more difficult. A trial-by-trial analysis of discrimination learning performance revealed that learning in the slow mode was abrupt, not incremental (Schmittmann, Visser, & Raijmakers, 2006). This result suggests that learning in the slow mode originated in inefficient hypothesis testing. More specific, Schmittmann and colleagues (2006) hypothesized that the inefficiency in learning is due to inefficient feedback processing in combination with inefficient hypothesis selection due to preferences for the irrelevant dimension. Berkeljon and Raijmakers (2007) investigated this hypothesis in a neural network model of the development of discrimination learning. The combination of a weaker influence of negative feedback (and, therefore, a relatively higher impact of positive feedback) and variability of the initial dimension preference resulted in two modes of output that resembled fast and slow abrupt learning. To further understand possible hypothesis testing strategies of children, it is useful to consider the following substrategies of efficient hypothesis testing (e.g., Dehaene & Changeux, 1991). A child using the win–stay substrategy randomly samples a rule from a set of rules and applies the rule until an error occurs. In addition, a child can use the lose–shift substrategy, meaning that a different rule is selected from the set of rules once an error is encountered. It is assumed that the new rule is sampled at random from the set. Studies of the hypothesis testing behavior of children suggest that young children might not apply the strict win–stay and lose–shift substrategies (Gholson et al., 1972, Kemler, 1978 and Phillips and Levine, 1975). However, these studies arrived at different conclusions, specifically concerning the use of the lose–shift substrategy in the age range of 4 to 10.5 years. This may be due in part to two indeterminacies. First, even if we observe lose–shift behavior on a given trial, we cannot conclude right away that a child actually applies a lose–shift substrategy. For instance, a child may forget to discard a falsified rule, so that the same rule may be sampled again on the subsequent trial, which is a lose–sample substrategy, or a child may use response stereotypes, which refer to position or stimulus feature-based responding that is insensitive to feedback (e.g., consistently choosing the left stimulus; Gholson et al., 1972). In the current study, we used mathematical modeling of the complete trial-by-trial data to distinguish between the underlying (e.g., lose–shift, lose–sample) substrategies. The second indeterminacy originates in the fact that children who choose the correct stimulus due to a marked preference for the correct feature do not need to process negative feedback, that is, apply a lose–shift or lose–sample substrategy. In this case, a win–stay substrategy is sufficient to reach the learning criterion quickly. Indeed, a win–stay strategy is not even necessary because a child with a marked preference can master the task simply by choosing the stimulus with the preferred feature, ignoring feedback altogether. Therefore, we have no means of knowing which substrategy a child who did not receive negative feedback actually applied. To solve this indeterminacy, we modified the task as described in detail below. Our modified task assesses feature preferences and forces all children who show a preference for a feature to learn a feature of their unpreferred dimension. Children without a preference may coincidentally start with the correct feature and should apply a win–stay substrategy to master the task. All remaining children should apply win–stay and lose–shift or lose–sample substrategies to master the task. The modification enables us to examine whether young children use efficient hypothesis testing strategies in discrimination learning. In addition, we can compare the learning processes of children who have a preference for a particular feature with the learning processes of children who have no such preference. Having a preference for an irrelevant feature might hinder the learning process if this preference is resistant or insensitive to negative feedback ( Kemler, 1978). A successful application of the win–stay and lose–shift substrategies requires working memory, attentional control, and cognitive flexibility, which are discussed in the following. Insufficient working memory resources, resulting in a failure to update and keep in mind the set of previously falsified hypotheses and to memorize the current hypothesis, may account for slow discrimination learning. When adults perform a distracter task with a high working memory load simultaneously with a discrimination learning task, their learning efficiency approaches that of children (in terms of the number of trials required to master the discrimination learning task; Sirois & Shultz, 1998). The possible role of working memory is supported by the finding that the dorsolateral prefrontal cortex and superior parietal cortex are involved in both feedback sensitivity (van Duijvenvoorde et al., 2008) and visuospatial working memory task performance (Klingberg, Forssberg, & Westerberg, 2002). A widely accepted conceptualization of working memory includes a supervisory system (the central executive), a phonologically based temporal storage system, and a visuospatially based temporal storage system ( Baddeley, 1992). Many working memory tasks, which are designed to measure working memory capacity in specific domains, are available. For example, tasks may require participants to update and manipulate verbal, numerical, spatial, and/or object information. In addition, tasks may vary in whether a repeated replacement of memory content or an accumulation of memory content is required. Besides working memory limitations, insufficient attentional control may contribute to slow discrimination learning. A successful learner needs to focus his or her attention on relevant stimulus information and to inhibit responses that are based on irrelevant stimulus information. Tasks such as the Stroop task (Stroop, 1935) and flanker task, in which participants are required to respond to the direction of a central arrow that is flanked by congruent or incongruent arrows (e.g., Fan et al., 2002 and Ridderinkhof and van der Molen, 1995), are thought to tax this ability. Cognitive flexibility seems to be particularly important to perform shifts between different hypotheses or dimensions in applying the lose–shift substrategy in discrimination learning (Ashby & O’Brien, 2005). The development of cognitive flexibility has been studied widely in a rule-following and rule-switching paradigm, based on the Dimensional Change Card Sort (DCCS) task (Zelazo, 2006) and DCCS modifications, which are appropriate for older children (e.g., Cepeda and Munakata, 2007, Deák, 2003 and Diamond and Kirkham, 2005). Studies based on the DCCS in its standard form show that a large proportion of children from 3 to 5 years of age are able to follow a sorting rule when informed of the rule. However, they perseverate on a rule that they sorted on previously when they are asked to switch to a rule that is based on a conflicting stimulus dimension (Zelazo et al., 2003). Different theoretical frameworks have been proposed to explain perseverative behavior on the DCCS task. According to the attentional inertia theory, children fail to suppress attention to the first dimension, such that they cannot shift attention to the second dimension (Kirkham, Cruess, & Diamond, 2003). In this theory, perseverating children fail to inhibit the prepotent responses that are associated with attention to the first (now irrelevant) dimension. Alternative explanations concern the ability to formulate higher order rules (cognitive complexity and control theory; Zelazo & Frye, 1997) and the relative strength of active memory (competing memory representations theory; Morton & Munakata, 2002). As mentioned above, our adapted discrimination learning task requires all children who show a preference for a feature to learn a feature of their unpreferred dimension and apply lose–shift or lose–sample behavior. An inability to do so leads to stereotypical responding or perseveration. The results reviewed above suggest the existence of different strategies of learning. However, the prevalence of the strategies in different age groups is unclear. In the current study, we addressed the development of different strategies in discrimination learning in 4- to 14-year-olds and examined whether strategy use was related to spatial working memory and attentional control. In addition, we investigated the effect of preferences for stimulus features on the learning performance. We hypothesized that young children who have a preference would have more difficulty in learning a feature of the unpreferred dimension than young children who have no such preference. With an increasing ability to switch between dimensions in older children, we expected this effect to wane with age.

Conclusion To our knowledge, this is the first study to investigate the role of the executive functions working memory and attentional control in different latent strategies of learning a simple discrimination of an unpreferred dimension. The current findings are consistent with the results of previous studies that revealed multiple learning modes in discrimination learning. A model with the three latent groups of hypothesis testing, slow learning, and nonlearning provided the best and most parsimonious fit to the data. The proportion of efficiently hypothesis testing children increased with age, and nonlearners were significantly present only in the 4- and 5-year-olds. The 4- and 5-year-olds who showed a feature preference and were forced to learn a rule of their unpreferred dimension appeared to be unable to engage in an efficient hypothesis testing strategy. However, a small group of the 4-and 5-year-olds who did not show a feature preference engaged in efficient hypothesis testing. Spatial working memory and attentional control significantly predicted posterior learning strategy probabilities after controlling for age.