Probabilistic category learning 1 RUNNING HEAD: Probabilistic category learning Challenging the Role of Implicit Processes in Probabilistic Category Learning

Considerable interest in the hypothesis that different cognitive tasks recruit qualitatively distinct processing systems has led to the proposal of separate explicit (declarative) and implicit (procedural) systems. A popular probabilistic category learning task known as the “Weather Prediction Task” is said to be ideally suited to examine this distinction because its two versions – ‘observation’ and ‘feedback’ – are claimed to recruit the declarative and procedural systems respectively. In two experiments we found results that were inconsistent with this interpretation. In Experiment 1 a concurrent memory task had a greater detrimental effect on the putatively implicitly mediated (feedback) version than on the explicit (observation) version. In Experiment 2 participants displayed comparable and accurate insight into the task and their judgment processes in both the feedback and observation versions. These findings have important implications for the study of probabilistic category learning in both normal and patient populations. Probabilistic category learning 3 In recent years there has been considerable interest in the hypothesis that different cognitive tasks recruit qualitatively distinct processing systems (e.g. Ashby, Alfonso-Reese, Turken & Waldron, 1998; Gabrieli, 1998; Knowlton, Squire & Gluck, 1994; Reber, Knowlton & Squire, 1996; Squire, 2004). The major distinction drawn by proponents of this multiplesystems view is between an explicit or declarative system that requires awareness and involves analytic processing, and an implicit or procedural system that operates in the absence of conscious awareness and is non-analytic. In addition to the distinction in terms of operational properties, some researchers suggest that specific neuroanatomical regions are differentially involved in mediating the two systems. One popular view is that the basal ganglia are involved in procedural learning, whereas the medial temporal lobes are involved in declarative learning (e.g. Ashby et al., 1998; Poldrack, Clark, Pare-Blagoev, Shohamy, Creso Moyano, & Myers, 2001; Shohamy, Myers, Onlaor & Gluck, 2004; Shohamy, Myers, Grossman, Sage, Gluck & Poldrack, 2004; Squire, 2004). A key piece of evidence for this proposed dissociation is the difference in performance between tasks that are learned via trial-by-trial feedback and those which are learned in an ‘observational’ manner with no feedback (e.g. Ashby, Maddox & Bohil, 2002; Poldrack et al., 2001; Shohamy et al., 2004a). Several authors have claimed that the procedural system is recruited when tasks are complex and learned incrementally via feedback. The feedback is crucial to engage the operation of a reward-related learning system thought to be mediated via dopamine neurons in the basal ganglia (Ashby et al., 1998; Schultz, Dayan, & Montague, 1997). Observation tasks are those in which, typically, stimuli are shown together with the correct outcome, no behavioral response is required and no feedback is provided. These tasks fail to engage the procedural system because there is no ‘surprising’ reward associated with learning. With such observational learning tasks, it is argued that performance on complex tasks is either impaired (Ashby et al., 2002) or a Probabilistic category learning 4 qualitatively different and neuroanatomically distinct system – (i.e. the declarative system) – takes over learning (Poldrack et al., 2001; Shohamy et al., 2004a) as the task requires a more pure form of memorization. Some of the evidence for these different patterns of performance comes from neuropsychological studies. Studies with Parkinson’s patients are of particular relevance because these patients suffer from a profound loss of dopamine-containing neurons in the substantia nigra. The loss of these neurons causes a decrease in striatal dopamine and disruption of basal ganglia function (Shohamy, Myers, Onalor, & Gluck, 2004). Filoteo, Maddox, Salmon, and Song (2004) demonstrated that Parkinson’s patients were impaired at learning complex ‘information integration’ tasks; the learning of which is thought to rely on the procedural system (though see Ashby, Ell, & Waldron, 2003). Similarly, Shohamy et al. (2004a) demonstrated that Parkinson’s patients were impaired in learning a feedback version of a probabilistic categorization task, but unimpaired, relative to controls, on an observation version of the same task. There are also several claims for differential involvement of the two systems in nonpatient groups. A study by Ashby et al. (2002) comparing performance on observation and feedback versions of information integration tasks showed impairments in the observation version in normal participants. Poldrack et al. (2001) demonstrated equivalent performance in normals on feedback and observation versions of a probabilistic categorization task but presented neuroimaging data suggesting differential engagement of the basal ganglia and the medial temporal lobe in the different versions. Presenting this dichotomous view of cognition, in which different tasks are claimed to recruit different systems, naturally invites simple experimental manipulations hypothesized to have differential effects on the two putative systems (see Ashby & Maddox, 2005 and Maddox & Ashby, 2004 for reviews of a comprehensive research program with this aim). In Probabilistic category learning 5 this paper, we draw on this simple strategy and scrutinize the processes underlying performance in a task that has been one of the primary sources of data for the proposed dissociation between the procedural and declarative system. The task is a prototypical probabilistic category learning task known as the weather prediction task (Knowlton et al., 1994). The task has been used in numerous investigations with a variety of populations (e.g. unimpaired individuals: Gluck, Shohamy & Myers, 2002; Lagnado, Newell, Kahan, & Shanks, 2005; Alzheimer’s patients: Eldridge, Masterman & Knowlton, 2002; Parkinson’s patients: Moody, Bookheimer, Vanek & Knowlton, 2004; Shohamy et al. 2004a; Shohamy et al. 2004b; amnesic individuals: Hopkins, Myers, Shohamy, Grossman & Gluck, 2004; Reber et al., 1996; and schizophrenics: Keri, Kelemen, Szekeres, Bagoczky, Erdelyi, Antal, Benedek & Janka, 2000) and has therefore been highly influential in the development of theoretical models of dissociable learning and memory systems (Squire, 2004). In the weather prediction task people learn to predict a binary outcome (rainy or fine weather) on the basis of four binary cues (four distinct tarot cards). Each card is associated with the outcome with a different probability, and these combine to determine the probability of the outcome for any particular pattern of cards. The trials presented in an experiment are made up of a representative sampling of possible card combinations. In the feedback version of the task, on each trial participants see a specific pattern of cards, predict the weather, and receive corrective feedback as to the correct outcome. This enables them to gradually learn the cue-outcome relations and thus improve the accuracy of their predictions. This feedback version is claimed to be mediated via the procedural system (Gluck et al., 2002; Knowlton et al., 1994; Poldrack et al., 2001; Reber et al., 1996; Shohamy et al., 2004a; Shohamy et al., 2004b). In the observation version of the task, on each trial participants are presented with both the cards and the outcome simultaneously. Crucially, this paired-associate arrangement Probabilistic category learning 6 does not rely on trial-by-trial feedback and so learning the task does not (perhaps cannot, cf. Ashby et al., 1998) recruit the procedural system. Rather, the declarative system is thought to be primarily responsible for learning (Poldrack et al., 2001; Shohamy et al., 2004a). Given this purported distinction between the two versions of the task, clear predictions can be made about the ways in which various experimental manipulations should affect performance in the observation and feedback versions. In Experiment 1 we make the simple prediction that if performance in the feedback version is mediated by an implicit procedural system (that is neuroanatomically distinct from the explicit system) then it should be less affected by placing additional demands on working memory. In contrast, the observation version, thought to be primarily under declarative control, should be detrimentally affected by an additional load on working memory. The logic of this manipulation is the same as that used by Waldron and Ashby (2001) in their investigation of the effects of concurrent working memory tasks on performance in information-integration and rule-based tasks. Although the feedback/observation distinction does not map directly on to the information-integration/rule-based one (see the General Discussion for elaboration of this point) the logic of the manipulation is the same. In their study Waldron and Ashby found that execution of a concurrent memory task (a numerical Stroop task) led to greater decrements in performance on a rule based task than an information-integration task. This finding lends support to their argument that the tasks are respectively mediated by explicit (working memory dependent) and implicit (working memory independent) systems. If we were to find such a pattern of results with the two versions of the weather prediction task then it would lend support to the idea that separate systems underlie performance in the feedback and observation versions. If on the other hand we were to find a general decrement across both tasks, or indeed a greater decrement in the feedback version Probabilistic category learning 7 than the observation version then this would question the current interpretation of performance in the weather prediction task. We should stress that part of our motivation for conducting the current experiments was recent research that has already questioned the characterization of the feedback and observation versions of the WEATHER PREDICTION task as being a priori procedural and declarative (e.g. Ashby & Maddox, 2005; Gluck et al., 2002; Lagnado et al., 2005; Shohamy et al., 2004a). Part of the problem with the task is that although it was designed as one that would be reliant on the procedural learning system, there are a number of ways in which a participant might attempt to ‘solve’ the task. Gluck et al. (2002) identified three types of strategies that appeared to account for the performance of most participants trying to solve the feedback version of the task. These were: (1) a multi-cue strategy in which participants learn about all four cards, and base their predictions on some integration of this information; (2) a singleton strategy in which they just learn about the cue patterns with a single card (and guess when more than one card is present); (3) a one-cue strategy in which they focus on just one card, and base their predictions on the presence or absence of this card. The structure of the task environment is such that these simpler strategies can be almost as successful as the more complex ones. Interestingly many participants in Gluck et al.’s experiments appeared to use the simpler, and potentially ‘explicit’, singleton strategy. However, Gluck et al. (2002) maintained that even this simple strategy could have been the product of an implicit system as the strategy identified through participants’ behavioral responses was typically unrelated to their verbal reports of the strategies they had adopted. Strategy analyses of the observation version of the task are as yet inconclusive, but Shohamy et al. (2004a) noted that aspects of their data from the observational version were inconsistent with the use of the explicit analytic strategies that would have been expected if the task was under declarative control. Probabilistic category learning 8 Perhaps even more importantly, Lagnado et al. (2005) recently provided strong evidence that the feedback version of the task involves explicit rather than implicit processes. Lagnado et al. (2005) elicited trial-by-trial ratings of participants’ reliance on each cue and demonstrated a clear divergence in the ratings given to weak and strong predictors of the outcome. This self-insight was coupled with excellent task knowledge demonstrated through an ability to judge accurately the probability of an outcome given the presence of a particular cue. In Experiment 2 we use these more sensitive and dynamic measures of insight to investigate the simple prediction that insight and task knowledge in the ‘declarative’ observation version should be superior to that in the ‘procedural’ feedback version. Again, if we find this to be the case it will lend support to the suggestion that the different versions of the task are mediated by different systems. If we find no difference or superior insight in the feedback version then again the current interpretation of the weather prediction task will be questioned. In the following experiments we capitalize on these recent advances in analyzing performance and insight in the weather prediction task. Our aim is to test the two simple predictions that arise from the multiple-systems view: 1) that placing additional demands on working memory should be more deleterious to performance in the observation version than the feedback version and 2) that insight and task knowledge should be superior following observation training than feedback training. Finding support for these two predictions would help to justify the extensive use of the weather prediction task in cognitive neuroscience as a ‘tool’ for demonstrating the operation of the hypothesized procedural and declarative systems. Finding results inconsistent with the predictions would suggest a re-evaluation of the claims made for performance in the weather prediction task. Probabilistic category learning 9 Overview of Experiments The experiments used the weather prediction task (WP) (Knowlton et al., 1994) employing both the observational and feedback versions of the task. In both versions of the task participants are shown between one and three discrete cues (easily discriminable tarot cards) on a trial-by-trial basis. In the feedback version participants are asked to predict a binary outcome (rainy or fine weather) and receive corrective feedback as to the actual outcome. In the observation version the outcome appears simultaneously with the cues and no overt prediction is required. In both experiments participants completed 102 training trials either under the observation or the feedback arrangement and then completed a further 42 test trials. The test trials required predictions but no corrective feedback was provided. These test trials served two purposes: 1) they allowed us to measure learning in the observation groups (note that in the observation group no overt prediction is required during training so no training phase data are recorded), 2) they allowed us to compare performance on the same sized sample of trials after equal of amounts training in the two versions of the task. This direct comparison has not been possible in previous research because of the failure to equate the number and type of test trials (e.g. Shohamy et al. 2004a). Experiment 1 compared performance in both versions of the WP task under either concurrent memory load or no concurrent memory load conditions. Experiment 2 did not employ a concurrent task but focused solely on the two versions of the WP task and sought more sensitive measures of both self and task insight. Probabilistic category learning 10