Connectionist, or Parallel Distributed Processing, mechanisms provide the foundation of most of my modeling efforts. This modeling framework offers a rich array of mathematical tools capable of both abstract characterizations of psychological processes and, with little or no augmentation, detailed accounts of neural events. The dynamic performance of such artificial neural network models can be quite complex, highlighting the need for both exploratory computer simulations and mathematical analysis to discover the entailments of detailed cognitive theories.
My initial investigations into rule-guided behavior have been restricted to one paradigmatic case: the enactment of explicitly provided instructions. Instruction following is a promising domain for the study of explicit rule use, since the explicit knowledge deployed in producing behavior is easily placed under experimental control. I have fabricated a general connectionist model of instruction following, and I have investigated the empirical validity of this model in the domain of instructed category learning.
Instructed category learning tasks involve the acquisition of a perceptual discrimination skill under the guidance of two sources of information: direct instructions in the form of explicitly provided categorization rules and experience with a collection of labeled examples. In typical instructed category learning situations, experience with examples tends to increase the automaticity of categorization, speeding responding and improving accuracy. This is not always the case, however. For example, experience with training examples can cause learners to deviate from explicitly provided instructions when classifying novel items, demonstrating a shift from rule-based categorization to a strategy dependent on similarity of the test item to previously seen training examples. Such exemplar-based interference is explained in my connectionist model in terms of the interaction between activation-based processing and weight-based error-correction learning. My model provides a good fit to the human data and also explains individual differences in this interference effect.
The human performance data against which this connectionist model was assessed arose from experiments which I performed to test predictions of the model. I continue to collect instructed category learning data from adult human learners, focusing on the influence of working memory load and on the impact of categorization rule complexity on the knowledge gained from training examples.
I have also conducted Bayesian optimality analyses of categorization and judgment tasks which involve explicit instructions. These investigations have shown how performance which violates provided instructions might be considered normative in the face of background assumptions concerning the reliability of explicit rules and the reliability of exemplar similarity as a predictor of category membership.
Instructed learning acts as good launching point from which a wide variety of other behavioral phenomena might be explored. For example, my connectionist model appears to be well suited to address instructional effects in strategy selection. The role of working memory in skill acquisition and the loss of verbal mediation as a skill becomes automatic may also be explored in the context of this model. Ultimately, I hope to augment this model to reflect mechanisms of hypothesis testing and planning, explaining how the rules that guide behavior might be generated from within rather than from an external teacher.
In hopes of understanding the neural basis of rule-guided and controlled behavior, I have embarked on a collaboration with Jonathan Cohen, Randy O'Reilly, and Todd Braver aimed at producing a robust computational model of DLPFC. Attractor network dynamics are used to actively maintain the contents of this working memory system, with the admission of new contents being primarily regulated by a dopaminergic signal from the ventral tegmental area. Following other efforts to model the function of midbrain dopamine neurons, we see this dopamine signal as encoding a "temporal-difference error" which may drive a process of reinforcement learning, shaping the system to retain only that information which is needed to acquire reward. Computational models employing this architecture have been used to explain behavioral and neuroimaging results involving delayed response tasks as well as the category learning behavior of monkeys with focal lesions. Current research has focussed on the learning of compositional representations in DLPFC over the course of development.
This collaboration has also brought to my attention the potential role of portions of anterior cingulate cortex (ACC) in the monitoring of cognitive conflict and the corresponding regulation of DLPFC contents. Recent work in the modeling of the ACC has inspired a modification to my model of instructed learning, in which the working memory system (DLPFC), maintaining information concerning explicit categorization instructions, is dynamically modulated by a conflict monitoring system (ACC) in response to task performance. It is hoped that this addition to the model will provide a biologically grounded mechanism for explaining decreases in working memory dependence as skills become more automatic.
![[--------------------------------------]](http://www.cnbc.cmu.edu/~noelle/share/spy_grad.gif){kind=small}