Prediction error and regularity detection underlie two dissociable mechanisms for computing the sense of agency

Abstract

The sense of agency refers to the subjective feeling of controlling one’s own actions, and through them, events in the outside world. According to computational motor control models, the prediction errors from comparison between the predicted sensory feedback and actual sensory feedback determine whether people feel agency over the corresponding outcome event, or not. This mechanism requires a model of the relation between action and outcome. However, in a novel environment, where this model has not yet been learned, the sense of agency must emerge during exploratory behaviours. In the present study, we designed a novel control detection task, in which participants explored the extent to which they could control the movement of three dots with a computer mouse, and then identified the dot that they felt they could control. Pre-recorded motions were applied for two dots, and the participants’ real-time motion only influenced one dot’s motion (i.e. the target dot). We disturbed participants’ control over the motion of the target dot in one of two ways. In one case, we applied a fixed angular bias transformation between participant’s movements and dot movements. In another condition, we mixed the participant’s current movement with replay of another movement, and used the resulting hybrid signal to drive visual dot position. The former intervention changes the match between motor action and visual outcome, but maintains a regular relation between the two. In contrast, the latter alters both matching and motor-visual correlation. Crucially, we carefully selected the strength of these two perturbations so that they caused the same magnitude of impairment of motor performance in a simple reaching task, suggesting that both interventions produced comparable prediction errors. However, we found the visuomotor transformation had much less effect on the ability to detect which dot was under one’s own control than did the nonlinear disturbance. This suggests a specific role of a correlation-like mechanism that detects ongoing visual-motor regularity in the human sense of agency. These regularity-detection mechanisms would remain intact under the linear, but not the nonlinear transformation. Human sense of agency may depend on monitoring ongoing motor-visual regularities, as well as on detecting prediction errors.

Introduction

People rapidly notice when environmental events depend on their actions. They then self-attribute these stimuli and feel sense of agency (Gallagher, 2000; Haggard & Chambon, 2012) over them. Thus, we readily perceive whether a sound of footsteps is caused by our own walking or not. Comparator models provide a framework examining this self-attribution process (Blakemore, Wolpert, & Frith, 2002; Blakemore, Wolpert, & Frith, 1998; Frith, Blakemore, & Wolpert, 2000; Wolpert & Flanagan, 2001). According to this computational framework, agency is computed by comparing a prediction generated from an internal model, using an efference copy of motor commands, and actual sensory feedback (blue background in Fig. 1). The sense of agency diminishes if the comparator produces a mismatch (i.e. a prediction error). For example, an incorrect internal model produces both poor motor control performance, and a large prediction error at the output of the comparator. In contrast, a well-adjusted internal model produces fluent motor control, together with predictable sensory inflow. This results in no prediction error at the output of the comparator, and, therefore, a “sense of agency”, or feeling that the action and feedback are self-caused. Besides the framework of comparator, many other motor control theories, such as the ideomotor theory (Prinz, 1997), suggest that the anticipation of action outcomes is important for motor control and the sense of agency (Spengler, von Cramon, & Brass, 2009). In other words, knowing what and how to control is considered to be a major premise of sense of agency.

On the other hand, active inference and learning theory suggests that behaviour has exploratory and exploitative aspects (Friston et al., 2016). Exploratory behaviours performed in environments containing ambiguity of control may rely on a cognitive mechanism that detects environmental statistics. Once the statistical relations between one’s exploratory actions and the events in the external world are detected, sense of agency may emerge. Once a sense of agency has been acquired through exploration, it can then be exploited in goal-directed behaviours, producing further evidence for self-attribution.

In this paper, we describe how extracting the natural statistics of motor and visual events during exploratory action involves a form of regularity detection. Importantly, regularity detection differs from model-based processes of motor control in that the former does not require a precise prediction of the outcome for each action. Developmental research offers some support for the idea of a regularity detection mechanism. Infants of 9–12 weeks infants’ made more foot thrust movements when their ankle was looped to an overhead suspension bar (Rovee & Rovee, 1969), so that their movements produced visual effects. Infants at this age have only minimal motor skill, and lack the precise forward and inverse models required for aimed movement. Therefore, the reinforcement of their exploratory behaviour is probably due to the perception of a regular relation between events in the external world (i.e. motion of the suspension bar) and their own actions (i.e. foot thrust). The perception of self-generated regularity can also contribute to sense of agency in later development. For example, this may account for the widespread preference for actions that reliably produce outcomes, over actions that do not (Karsh, Eitam, Mark, & Higgins, 2016; Nafcha, Higgins, & Eitam, 2016).

In the present study, we investigate whether motor performance and sense of agency are inextricably linked, as comparator models suggest, or may dissociate. On the former view, an efficient internal model leads to both good motor performance and strong sense of agency, as a result of a common cause of reduced low prediction errors. On another view, different computations might underlie objective motor performance and subjective sense of agency, potentially producing dissociations. Here we investigate a possibility that sense of agency is partly based on detecting global regularities between one’s actions and their outcomes through repeated sampling (pink background in Fig. 1). This process would bypass the internal (forward and inverse) model (Kawato, 1999; Wolpert & Kawato, 1998), using buffered samples of action and sensory feedbacks to calculate a correlation between them. Importantly, this putative regularity mechanism would detect a sense of agency using the same kind of pattern-detection processes used in other cognitive and perceptual functions, and without any special or privileged link to the motor control system that decides and generates actions. When we perform a series of actions successfully, the comparator mechanism repeatedly signals no prediction error, and the regularity mechanism accumulates repeated samples in which action and outcome are strongly related. However, the two mechanisms are potentially dissociable. For example, across a series of actions, there might be a consistent prediction error, perhaps due to an unadapted internal model, yet there might be a regular relation between actions and outcomes. For example, action and sensory feedback in Fig. 1 show a strong correlation but a consistent non-zero prediction error (appearing as a positive intercept in the scattergram in lower panel of Fig. 1). In this case, the comparator mechanism would suggest no sense of agency should occur, while a regularity mechanism would suggest that normal sense of agency should be present. In addition, higher-level beliefs in one’s own agency may also affect the processes in motor control and regularity detection, although this notion was not shown in Fig. 1.

To test the dual-route hypothesis of sense of agency shown schematically in Fig. 1, we induced two factors that disturb the sense of agency by selectively influencing the prediction-error mechanism and the regularity-monitoring mechanism respectively. First, we provided a constant disturbance by applying a consistent angular bias in the visual feedback caused by participants’ actions. This leads to large prediction errors at the comparator, but leaves intact any regularity in the relation between action and feedback. Given sufficient experience, people readily adapt to such consistent disturbances by updating their internal model, and prediction error decreases rapidly (Kitazawa, Kimura, & Uka, 1997). However, in our task, we minimised such adaptation by randomly mixing trials with different angular biases (e.g. 30° and 90°). As a result, a non-zero prediction error persisted, even though the relation between action and visual feedback was highly regular. The second form of disturbance involved mixing 60% of someone else’s motion with the participants’ own instantaneous motion in order to generate the visual feedback (Wen & Haggard, 2018; Wen, Brann, Di Costa, & Haggard, 2018). The highly nonlinear disturbance introduces errors at the comparator, which vary from one moment to the next. Importantly, such nonlinear disturbance produces a highly irregular relation between action and visual feedback. Note that the nonlinear and linear disturbances were alternatives, and were never combined in a single trial. In other words, participants either experienced an angular transformation or a mixture of their motion and someone else’s motion in separate conditions. Finally, we ensured that these two qualitatively-different disturbances had quantitatively similar effects on performance in our motor control task (see below). This meant we could investigate the extent to which each kind disturbance might influence the sense of agency, without confounding differences in motor performance. In addition, we also included a condition of 30° angular bias, and a condition of 90% other’s motion, for references of good and poor control, respectively. Therefore, we measured both motor control performance and the sense of agency, using a reaching task and a control detection task, respectively, in each of the above four disturbance conditions (30° and 90° angular bias, 60% and 90% others’ motion).