Motor Dynamics of Task Switching Nicholas C. Hindy 1 (hindy@psych.upenn.edu) Michael J. Spivey 2 (spivey@cornell.edu) 1 Department of Psychology, University of Pennsylvania, Philadelphia, PA, 19104 2 Department of Psychology, Cornell University, Ithaca, NY, 14853 Abstract The traditional view of cognitive-motor control posits an executive module that sends instructions to motor cortex before each movement. In contrast, newer work suggests that manual movements in task switching are determined through the competition of multiple simultaneously active motor representations. The streaming x,y coordinates of participants’ computer mouse movements were recorded as they switched between conflicting or compatible sorting tasks. Through analysis of the mouse-movement trajectories, we evaluate these theoretical perspectives. Although switching from the more difficult to the less difficult or vice versa resulted in comparable overall response time cost, the movement associated with these two types of task-switches exhibited quite different motor dynamics. Results are interpreted within a dynamical systems perspective on cognitive control, in which graded motor representations compete with one another during reaching movements after a task-switch. Keywords: Psychology; motor control; attention Introduction Cognitive control is the ability to override impulses, ignore distractions, and intentionally guide thoughts and actions in the pursuit of internal goals. Cognitive control is essential to task switching and multitasking, and is most crucial when the same signal may lead to different potential appropriate responses, depending on the situation. Automatic behaviors, whether innate or learned over time, are inflexible and stereotyped. Because such actions are elicited by stimuli in the outside environment, they are usually referred to as “bottom-up” behaviors. Conversely, “top- down” behaviors are based on internal goals or intentions. Through cognitive control, top-down signals attempt to override the automatic bottom-up behaviors when they are inappropriate. During most experiments concerning cognitive control, participants are requested to inhibit a certain type of information at a particular stage in their cognitive processing. Zelazo, Frye, and Rapus (1996) introduced the dimensional-change card-sort (DCCS) task to test cognitive control development in young children. A set of cards includes a blue truck, a red truck, a blue star, and a red star. Participants are initially asked to sort the test cards by one of the dimensions, either color or shape. Then after several trials, they are asked to switch to the other sorting dimension, and sort the same stimuli according to the new dimension. For example, when sorting by shape, the red truck and blue truck go together, but when sorting by color, the red truck and red star go together. Diamond and Kirkham (2005) adapted the DCCS task for use on a computer, and found that if adult participants first sorted these images by color, their reaction times would increase by more than 300 ms when they switched to sorting by shape (or vice versa). Two types of models have been proposed to account for the time lags found in task-switching experiments: stage- based models of task-set reconfiguration (Rogers & Monsell, 1995; Kieras, Meyer, Ballas, & Lauber, 2000; Meiran, 2000), and continuous dynamic models of interference resolution (Allport & Wylie, 2000; Gilbert & Shallice, 2002). The goal of the present study is to distinguish between these two classes of models by looking at the distributions of participants’ computer mouse curvatures as they switch between sorting dimensions in a computer version of the DCCS task. Most stage-based reconfiguration models of task switching contend that delayed reaction times after a task- switch are due to participants’ need to endogenously reconfigure connections between sensory and motor modules before they can adopt a different task-set. These models are usually implemented as fixed cognitive architectures with control operations and adjustable parameters at each processing stage. For instance, Kieras et al.’s (2000) EPIC model includes a rule-based “central cognitive processor” and mechanical effectors. To simulate human performance in a choice reaction time task-switching experiment, the EPIC model attributes increased latency effects to the additional time it takes the cognitive processor to reconfigure for the new task. Once this cognitive reconfiguration is complete, the cognitive processor sends its command to the motor processors. Through adjustment of time and processing parameters, Kieras et al. fit an otherwise fixed model to numerous reaction time datasets. Most interference resolution models of task switching contend that the time-cost after a task switch is due to the “carryover effect” of persisting activation of the previous task, as well as associative learning between the stimuli and the task. Interference models have been implemented as parallel distributed networks of interconnected nodes. Gilbert and Shallice (2002) used an interactive activation model to simulate human performance when switching between word- reading and color-naming in the Stroop task. Their model had separate pathways for word reading and color naming, as well as a “task demand” input to bias the network toward the relevant task. To simulate the asymmetric interference between word reading and color naming, the connection 2474