Multiple Systems of Category Learning

Divergent theoretical perspectives on category learning have been hotly debated for more than 50 years. Perhaps the most important finding to emerge from this debate is that no one theoretical perspective can account for category learning. Instead, it appears that people have multiple systems that are capable of category learning and learning in general. Critically, the system(s) that mediate learning to depend upon a complex combination of task demands, internal motivation, external pressure, and neurological health (among other factors).

My work to date has focused on characterizing two category learning systems: hypothesis-testing and procedural-based learning systems. The hypothesis-testing system is constrained to use logical rules that are applied to make categorization decisions. At the cognitive level, rule learning depends critically upon the ability to maintain the current rule in working memory, retrieve previous rules from long-term memory, the selection of novel rules, and the ability to switch executive attention among competing rules. At the neural level, the hypothesis-testing system depends primarily upon a network of cortical (lateral prefrontal cortex, premotor cortex, anterior cingulate) and subcortical (basal ganglia, thalamus, and midbrain dopaminergic cells) structures.

The procedural-based system learns by incrementally associating stimuli with a particular categorization response. In contrast to the hypothesis-testing system, the procedural-based system does not have a strong dependence upon the maintenance and manipulation of categorization rules and, as a result, is not as dependent upon working memory. Instead, the procedural-based system is critically dependent upon trial-by-trial, corrective feedback to strengthen the appropriate stimulus-response associations. At the neural level, the procedural-based system depends upon a network of cortical (inferotemporal cortex, supplementary motor area) and subcortical (basal ganglia, thalamus, and midbrain dopaminergic cells) structures that are distinct from the hypothesis-testing system.

In collaboration with researchers at UC Santa Barbara (Greg Ashby), UC Berkeley (Rich Ivry), and UT Austin (Todd Maddox), I have been able to dissociate performance in tasks designed to be probes for the hypothesis-testing and procedural-based systems (Ashby & Ell, 2001, 2002a, 2002b; Ashby, Ell, & Waldron, 2003; Ell, Cosley, & McCoy, 2011; Ell, Hutchinson, & Maddox, 2012; Ell & Zilioli, 2012).  For example, I demonstrated that category separability (i.e., d′) can be used as an effective means to influence which system is in control (Ell & Ashby, 2006). I have also done considerable work on the neural substrates of these category learning systems using rigorous behavioral methodology to test individuals with focal basal ganglia lesion due to stroke (Ell, Marchant, & Ivry, 2006; Ell, Weinstein, & Ivry, 2010), cerebellar pathology (Ell & Ivry, 2008), ventromedial prefrontal lesions (Schnyer et al., 2009), and Parkinson’s disease (Ashby, Noble, Filoteo, Waldron, & Ell, 2003; Ell, et al., 2010). Importantly, the primary goal of this neuropsychological work is not simply to investigate the correspondence between neural systems and category learning systems, but rather to investigate the specific cognitive processes that are affected (e.g., decision making in the presence of distracting information, rule maintenance, and updating).

The Stress-Cognition Relationship

The way in which we respond to everyday stressors can have a profound impact on cognitive functioning. Maladaptive stress responses, in particular, are generally associated with impaired cognitive performance. I argue, however, that the cognitive system mediating task performance is also a critical determinant of the stress-cognition relationship. In this line of work, I combine techniques from Social Psychology, Psychophysiology, Cognitive Psychology, and Cognitive Neuroscience in an effort to understand how variability in the stress response interacts with variability in the cognitive system mediating task performance to affect cognitive functioning.

In a recent collaboration with Shannon McCoy (Ell, et al., 2011), I investigated this question by examining the impact of a social-evaluative stressor that reliably induces a physiological stress (i.e., the Trier Social Stress Test – Kirschbaum, Pirke, & Hellhammer, 1993) on subsequent category learning performance. In contrast to previous research, I predicted that the type of stress response (not the mere presence of a stress response) would be critical in determining the impact on categorization performance and that this effect would vary depending upon the category learning system that was mediating performance. Consistent with this prediction, I found that physiological and psychological stress reactivity consistent with a maladaptive, threatening stress response predicted enhanced performance for the procedural-based system and a trend for impaired performance for the hypothesis-testing system. This represents one of the first published demonstrations that a maladaptive, threat-related stress response is associated with enhanced cognitive performance on an emotionally-neutral task.

A closely related question in Cognitive Psychology centers on how pressure impacts cognitive performance. Indeed, many researchers have gone so far as to argue that pressure is stressful. In collaboration with Shannon McCoy, I tested this prediction by investigating the impact of pressure on physiological measures of stress reactivity during category learning. Consistent with previous assumptions in the pressure literature, pressure elicited a modest physiological stress response. The stress response to pressure, however, differs from the stress response to social-evaluative stress in two important ways. First, the stress response to pressure is considerably diminished. Second, the stress response to pressure is not predictive of category learning performance. A manuscript reporting these results is in preparation (McCoy, Hutchinson, Hawthorne, Cosley, & Ell, 2012).

Rule-Guided Behavior

In the most general sense, rule-guided behavior refers to a conditional mapping between some input (e.g., a stimulus, an internal representation) and some output (e.g., a behavioral response, another internal representation) (e.g., Bunge & Wallis, 2007). Despite the apparent simplicity of this process and the ease with which we use rules to guide our behavior, a comprehensive understanding of rule-guided behavior remains elusive. Using rule-based categorization as a model task, I am investigating some of the key components of rule-guided behavior (e.g., working memory, selective attention, rule implementation, and rule switching).

In many rule-based categorization tasks, participants must learn which features are relevant for successful performance and apply rules on the relevant features. In a recent collaboration with Todd Maddox (Ell, Ing, & Maddox, 2009), I investigated the impact of interfering with working memory processes necessary for rule maintenance and updating. Based on predictions generated from Ratcliff’s diffusion drift model (Ratcliff, 1978), we were able to demonstrate that increasing working memory demand impaired rule maintenance and updating. We also demonstrated that increasing the number of relevant features attenuates this impairment – an extension of a classic finding from the absolute identification literature (e.g., Miller, 1956).

The vast majority of category learning theories have focused on supervised category learning (i.e., the ability to learn categories with the aid of corrective feedback). Several recent theories, however, have incorporated mechanisms for unsupervised category learning (i.e., the ability to learn categories without the aid of corrective feedback). Most empirical research on unsupervised categorization has focused on the question of how participants prefer to construct categories. Equally important is the question of what types of categories participants are capable of learning under unsupervised conditions. I have investigated this latter question in a series of papers focusing on the limitations of the unsupervised learning of rule-based categories (Ell & Ashby, 2012; Ell, Ashby, & Hutchinson, 2012).

In addition to my more traditional cognitive work, I have investigated the neural substrates of rule-based categorization using a combination of neurocomputational and neuropsychological techniques. In collaboration with Greg Ashby, I developed a neurocomputational model of working memory maintenance that successfully accounted for single-cell recording data as well as human behavioral data (Ashby, Ell, Valentin, & Casale, 2005). My neuropsychological work with individuals with Parkinson’s disease (PD – a neurodegenerative disease that disrupts processing in prefrontal-basal ganglia neural networks) suggests that the rule maintenance and updating necessary for learning rule-based tasks is dependent upon intact prefrontal-basal ganglia function (Ashby, Noble, et al., 2003; Ell, et al., 2010).

Although much has been learned about prefrontal-basal ganglia contributions to rule-guided behavior from the study of individuals with PD, very few attempts have been made to adapt experimental techniques to test behavioral training paradigms for improving rule-guided behavior in PD patients. Moreover, previous attempts to improve the ability of PD patients to cope with the cognitive symptoms of the disease have used fairly coarse training protocols, making it difficult to determine which aspects of training were critical for improvement and what cognitive processes were affected. In a paper I recently submitted for publication, I addressed this challenge by demonstrating that targeted training of the decision rule improved performance on a rule-based categorization task. Importantly, I also demonstrated that the benefit of targeted training transferred to a subsequent test of rule-guided behavior (i.e., the Wisconsin Card Sorting Test) (Ell, 2012a).