Relative roles of appetitive-reward and aversive-punishment reinforcement in acquiring categories and their underlying and interacting neurobiology

Student: Vivian Valentin
Current Lab: Greg Ashby (UC, Santa Barbara)
Host Lab: Alan Pickering (Goldsmiths College, London)

The goal of this research is to develop a neurobiologically plausible model of feedback processing during category learning. Such a model needs to include a mechanism that learns from punishment, a mechanism that learns from reward, and some mechanism that can integrate the resulting motivational outputs of the two systems. A neuropsychological theory developed in Ashby's lab (Ashby et al., 1999; Ashby, Valentin, & Turken, 2002) predicts that dopamine will be released from the ventral tegmental area following unexpected positive feedback. The theory predicts that this in turn should enhance hypothesis generation (during category learning), which is carried out by a network that includes the prefrontal cortex and anterior cingulate. Dopamine is known to modulate the effects of glutamate in this circuit. Ashby and Casale (2003) derived a computational model of this neuromodulation that assumes dopamine potentiates the glutamate response through the NMDA receptor and it depresses the glutamate response through non-NMDA receptors (e.g., AMPA). The model successfully accounts for human in vitro single-cell recording data that examined the effects of dopamine on the firing rate of glutamatergic cortical cells. Pickering has worked extensively on the punishment system, and its effects in procedural, associative, and instrumental learning (e.g., Corr, Pickering, & Gray, 1997). Briefly, this model proposes a punishment network that includes interactions among the amygdala, hippocampus, prefrontal cortex, and orbitofrontal cortex (among other structures). In addition, an inhibitory influence of D2-dopamine receptors (as included in previous modeling work by Salum, Roque-da-Silva, & Pickering, 1999) may operate at other ventral striatal neurons that receive inputs from the amygdala punishment system. We will explore the possibility that these two systems interact via a control system that includes the orbitotfrontal cortex and the ventral striatum. The resulting model will make predictions about the differential effects of positive and negative feedback in a variety of different types of category learning tasks (e.g., rule-based, information-integration).

==============================================================

Collaborative Study of Category Learning in Amnesia

Student: Julia High
Current Lab: Tom Palmeri (Vanderbilt)
Host Lab: Mona Hopkins (BYU/Univ. of Utah)

A fundamental question for the cognitive neuroscience of category learning is whether amnesics learn novel categories as well as normal individuals. Studies by Squire, Knowlton, and colleagues and by Filoteo, Maddox, and colleagues have reported preserved category learning by amnesics in the face of significant deficits on a variety of other memory tasks. Such dissociations suggest that category learning is subserved by a functionally independent memory system. However, Nosofsky and Zaki and Shanks and colleagues have demonstrated that many of these behavioral dissociations are entirely consistent with the predictions of single memory-system models. Furthermore, Palmeri and colleagues have shown that many of the category learning paradigms used to test amnesics have fundamental methodological flaws. The proposed project will first develop new category learning tasks that do not suffer from the methodological problems of previously used tasks. Julia High will assist with the pilot testing of these new tasks. The proposed project with next test amnesics and normal controls on these methodologically improved tasks. Julia High will assist with testing amnesics. The question is simple: Will amnesics learn novel categories as well as normal individuals when improved category learning tasks are used? The focus of this project is a study by Filoteo, Maddox, and Davis (2001), who investigated whether amnesics could learn to classify stimuli defined by a "complex categorization rule." Adapting a well-known paradigm developed by Ashby and colleagues, subjects learned two categories defined by multivariate normal distributions. On every trial, one of the two categories (normal distributions) is randomly selected, a stimulus is sampled from the category distribution, the subject classifies it as a member of Category A or Category B, and receives corrective feedback. The figure below displays contours of equal likelihood for the categories used in the Filoteo et al. study. Stimuli were composed of a horizontal and vertical line connected at the top left corner and varied in the length of the horizontal and vertical components. Filoteo et al. reported that amnesics learned these categories as well as normal controls. In the language of general recognition theory, they claimed that amnesics can learn categories defined by a "complex quadratic rule" that requires integrating information across two dimensions.

However, Palmeri and Flanery (2002) noted that Category A stimuli have equal height and width ("squares") and Category B stimuli vary in height and width ("rectangles"). In other words, these stimuli can be alternatively described in terms of shape and area - Category A stimuli are "squares" varying in size and Category B stimuli are "rectangles" varying in size. Although Filoteo et al. described these categories in terms of learning a complex quadratic rule, an alternative interpretation is that amnesics were able to learn an exceedingly simple shape rule ("square" versus "rectangle"). Indeed, the fact that both amnesics and normals reached near asymptotic performance after only a small number of trials supports our claim that the task was learned using a very simple rule - other research by Ashby and Maddox found that learning categories defined by quadratic rules often require multiple sessions to reach asymptote. Moreover, Palmeri and Flanery replicated and extended the Filoteo et al. task but instead used stimuli that did not permit the use of emergent dimensions. When stimuli were circles that varied in size that had a diameter line that varied in angle, all but one of five subjects failed to exceed chance performance after the first block of 200 trials, and three of five subjects failed to exceed chance performance after the entire set of 600 trials. The ability to learn a simple shape rule is rather weak evidence for an entirely preserved category learning system in amnesia. We aim to test category learning in amnesics using stimuli and category structures that do not permit the formation of a very simple rule. The most straightforward response to Filoteo et al. would be to replicate their study using circle-line stimuli rather than line-line stimuli. But if motivated graduate students cannot learn categories composed of the circle-line stimuli then we certainly cannot expect elderly brain-damaged individuals to be able to learn categories composed of the circle-line stimuli.
So, our first aim is to construct a new category structure that retains the essential character of the Filoteo et al. structure but that can be learned using circle-line stimuli (stimuli that do not give rise to emergent dimensions that permit for the formation of a simple rule). The goal of initial pilot testing will be to test whether young adults at Vanderbilt and Indiana can learn the new category structures. The group at Rutgers will next test whether elderly controls can learn the new category structures. Julia High will then assist Hopkins in Utah in testing amnesics on the category structure that emerges from the pilot testing. Both amnesics and controls will learn the categories using line-line stimuli and will learn the categories using circle-line stimuli. Adapting the Filoteo et al. study, subject will be given 200-300 category learning trials with feedback. We expect to replicate the Filoteo et al. results using the line-line stimuli - that amnesics can learn a simple categorization rule as well as normal controls. The important new question is whether amnesics will be able to learn the circle-line stimuli as well as normal controls, as predicted by Filoteo et al. We will also test transfer of category knowledge after a brief delay, as was done by Filoteo et al. But we will test for transfer using more diagnostic measures of category knowledge, as proposed by Palmeri and Flanery (2002).

==============================================================

fMRI of Information-integration and Rule-based Categorization

Student: T.B.A
Current Lab: Vince Filoteo (UCSD) & Todd Maddox (Texas)
Host Lab: Paul Reber

(Northwestern)Maddox, Ashby & Bohill (in press) recently demonstrated a dissociation between information-integration (II) and rule-based (RB) categorization by manipulating the delay before feedback was given. During II categorization, learning was disrupted when there was a delay between the participant's response and feedback, while rule-based categorization was not affected by the delay. The dissociation is hypothesized to arise from the different neural systems thought to support the two types of category learning. The behavioral result appears to parallel recent studies that have used fMRI to examine the neural basis of category learning (Poldrack et al. 2001; Reber et al. in press). We propose to examine the brain activity evoked during II and RB categorization using fMRI. The categorization task will be adapted for an event-related fMRI design collaboratively between Todd Maddox (Univ. Texas at Austin), Vince Filoteo (UCSD), and Paul Reber (Northwestern). The data collection site will be the research-dedicated 3T Siemens Trio MRI at Northwestern University. In addition to comparing brain activity associated with the two types of category learning, we also anticipate comparing (1) activity during successful and unsuccessful categorization, (2) changes in activity over the course of learning, and (3) differential activity for the two categories learned. This collaborative project should take maximal advantage of the expertise of the three primary investigators in the behavioral methodology of this categorization task (Maddox), previous neuropsychological categorization studies (Filoteo) and functional neuroimaging (Reber).

==============================================================

Basal Ganglia Contributions to Category Learning: The Effect of Deep Brain Stimulation on Leaning in Parkinson's Disease

Student: Daphna Shohamy
Current Lab: Mark Gluck (Rutgers)
Host Lab: David Eidelberg (North Shore Long Island Jewish Hospital)

Recent studies have emphasized a critical role for the basal ganglia in probabilistic category learning. For example, neuorimaging studies with healthy controls have demonstrated that probabilistic category learning is associated with increased activation in basal ganglia nuclei. In Parkinson's disease (PD), individuals suffer from disrupted basal ganglia function due to a loss of dopaminergic projections from the substantia nigra compacta. Several studies have now demonstrated that PD patients are impaired at probabilistic category learning on a number of paradigms and with different kinds of stimuli (e.g. Knowlton et al., 1996; Shohamy et al., 2003a,b). In some cases, motor symptoms of PD are treated with deep brain stimulation (DBS), which is effective in alleviating motor symptoms in many severely affected patients. Recent studies indicate that DBS may also affect cognitive function in PD patients. Studies by Eidelberg and colleagues have been investigating the effects of DBS on motor and learning behavior in PD patients (e.g. Carbon & Eidelberg, 2002; Fukuda et al., 2002). They found that DBS improved motor sequence learning (Fukuda et al., 2002). DBS was also associated with enhanced learning related activity in basal ganglia nuclei (Fukuda et al., 2002). Here, we propose to investigate the effect of DBS on category learning in patients with PD. PD patients will be tested either 'on' or 'off' DBS, using a battery of category learning tasks which have previously been shown to be sensitive to basal ganglia function. We expect these studies to provide important insights into the neural bases of category learning, and to allow more specificity in understanding how different subregions of basal ganglia nuclei contribute to category learning.