Mechanisms of skill refinement : A model of long-term repetition priming

We address an omnipresent and pervasive form of human learning— skill refinement , the improvement in performance of a cognitive or motor skill with practice. A simple example of skill refinement is the psychological phenomenon of long-term repetition priming : Participants asked to read briefly presented words are more accurate if they viewed the word earlier in the experiment. We simulate various phenomena of repetition priming using a probabilistic model that describes information flow along cortical processing pathways. The model suggests two distinct mechanisms of adaptation with experience, one that updates prior probabilities of pathway outputs, and one that improves information transmission through a pathway. These two mechanisms loosely correspond to bias and sensitivity changes that have been observed in experimental studies of priming. Both mechanisms are extremely sensible from a rational perspective, and serve as the foundation of skill acquisition and skilled performance. Computational modeling has focused primarily on two aspects of human learning—the induction of new concepts and categories, and the acquisition of new skills. Another aspect of human learning has received little attention—the refinement of existing skill. Skill refinement, also called skill practice, is an omnipresent and pervasive form of learning. As we type, drive, read, or play video games, our behavior becomes less error prone and more fluent, rapid, and robust to distraction and irrelevant aspects of the task. Skill refinement is sometimes explicit, such as the rehearsal of a piano sonata, but is often implicit, such as entering one’s personal identification number at an automated teller machine. Understanding skill refinement is fundamentally about discovering the mechanisms by which one trial or performance of the skill leads to improvements on the next trial. 1 Long-term repetition priming Perhaps the most direct and easily studied manifestation of skill refinement in the psychological literature is the phenomenon of l ng-term repetition priming . In the priming paradigm, participants engage in a series of experimental trials, and experience with a stimulus or response on one trial results in more efficient processing on subsequent trials. Efficiency is defined in terms of shorter shorter response times, lower error rates, or both. A typical long-term perceptual priming experiment consists of a study phasein which participants are asked to read a list of words one at a time, and a test phase , during which they must respond a series of brief, masked target words. The time between target onset and mask onset is is called the flash duration. Typical response paradigms include speaking the target aloud ( naming) and a forced choice between two alternatives ( 2AFC). Repetition priming occurs when a word from the study phase influences performance during the test phase. Priming is an implicit memory phenomenon: participants are not told the study and test phases are related, and they do not try to recall study words during the test phase as a deliberate strategy for performing the task. Thus, priming is incidental and not task related; it comes about as a result of experience and is thus a form of skill refinement, where the “skill” here is perceptual processing of a letter string. Priming in this paradigm is long term, in that it persists over a period of many minutes and many intervening trials. Priming can also be short term, in that it persists only from one trial to the next. Priming can occur based not only on repetitions of an item, but based on semantic or orthographic similarity. Repetition priming is an easy case to study because it is the case of maximal similarity between prime and target. Models have been proposed for other forms of priming (e.g., Huber et al., 2001). A key question concerning repetition priming is whether—to use the language of signal detection theory—priming is due to increased biasor increasedsensitivity. Bias means that participants are more likely to report studied items regardless of what word is presented for identification. Sensitivity means that participants become better at perceptual discrimination of the studied items. From signal detection theory, it is well known that an increased bias toward a studied word can either benefit (by increasing the correct detection rate) or hinder (by increasing the false detection rate) overall performance. In contrast, increased sensitivity to a studied word has the specific effect of improving the ability to perceive that word during the test phase. A key finding in long-term repetition priming research has been that priming reflects both increased bias and increased sensitivity, although the sensitivity increase is robust only for low-frequency words or novel items. The goal of this paper is to introduce a model of skill rehearsal and performance. The model has two distinct learning mechanisms which contribute to skill improvement with practice. The model explains various data from psychological studies of long-term repetition priming. In this paper, we model two experiments isolating bias and sensitivity effects in priming, and show that our two learning mechanisms correspond to these two effects. 2 Modeling long-term repetition priming Our theory posits that cortical computation is performed by a set of functionally specialized pathways. Each pathway performs a primitive cognitive operation, e.g., visual word-form recognition, identification of semantic features of visual objects, computation of spatial relationships, or construction of motor plans. To model the effects of long-term repetition priming, we propose a model with two pathways in cascade. A perceptualpathway maps visual features to word identities. A responsepathway takes the output of the perceptual pathway and maps it to a task-appropriate response. We assume the pathways communicate continuously during processing and that communication is unidirectional. 2.1 Implementing a pathway as a dynamic belief network We present a probabilistic model of a pathway, which characterizes the transformation of pathway input to output, the time course of information processing during a single trial, and tuning of the pathway behavior over many trials. The inputs and outputs of a pathway (a) 0 0.2 0.4 0.6 0.8 1