The perceptual shaping of anticipatory actions

Abstract

Humans display anticipatory motor responses to minimize the adverse effects of predictable perturbations. A widely accepted explanation for this behaviour relies on the notion of an inverse model that, learning from motor errors, anticipates corrective responses. Here, we propose and validate the alternative hypothesis that anticipatory control can be realized through a cascade of purely sensory predictions that drive the motor system, reflecting the causal sequence of the perceptual events preceding the error. We compare both hypotheses in a simulated anticipatory postural adjustment task. We observe that adaptation in the sensory domain, but not in the motor one, supports the robust and generalizable anticipatory control characteristic of biological systems. Our proposal unites the neurobiology of the cerebellum with the theory of active inference and provides a concrete implementation of its core tenets with great relevance both to our understanding of biological control systems and, possibly, to their emulation in complex artefacts.

Keywords: active inference; cerebellum; computational model; motor control; perceptual learning.

Figure 1.

Conceptualization of the Hierarchical Sensory Predictive Control (HSPC) hypothesis. A predictable displacement caused by a soccer ball directed to the chest elicits an anticipatory response that minimizes the loss of balance before it is perceived. In HSPC, the anticipatory response is the result of a hierarchy of descending sensory predictions from distal (visual detection) to proprioceptive (impact) to vestibular (loss of balance) modalities, where each modality advances in time the expected consequences on the next modality until the predicted error in balance triggers a reflexive action in a feed-forward manner. The minimal model for this behaviour is an inverted pendulum of mass (m) and height (h), whose error in angle (θ) is minimized by generating a torque (τ) at the ankles that counteracts the disturbance (F).

Figure 2.

Motor anticipation (FEL) versus sensory prediction (HSPC) strategies. (a) Different responses are elicited by different sensory modalities. (i) A corrective reaction is triggered by the perceived postural error. (ii) A fast compensatory corrective action is triggered by the perceived impact (proximal stimulus). (iii) An anticipatory action is triggered by the distance to the obstacle (distal stimulus). (b) Motor anticipation strategy (FEL). (i) A postural error is converted into a reflexive action by a feedback controller (R). (ii) A feed-forward compensatory action associated with the impact signal is acquired by the proximal adaptive module (FFp) on the basis of the feedback response to the error. (iii) A feed-forward anticipatory action associated with the distal cue is acquired by the cerebellar distal module (FFd) on the basis of the same feedback response. (c) Sensory prediction strategy (HSPC). (i) Reflexive action elicited as in FEL. (ii) Feed-forward compensatory action: triggered by the proximal cue and learned from the closed-loop error, a counterfactual error is issued by the proximal module (FFp) in response to the proximal cue driving the feedback controller. (iii) Anticipatory action: evoked by the cue, a prediction of the expected impact issued by the distal module (FFd) triggers the compensatory response in an anticipatory manner.

Figure 3.

Experimental results. Acquisition. (a) Mean angular position during the disturbance rejection task for feedback-control condition (grey – 10 trials), trained FEL architecture (red – trials 90–100) and trained HSPC architecture (cyan – trials 90–100). Disturbance is delivered at t = 0 (dashed line). (b) Root mean square error (RMSE) in angular position over trials during acquisition phase for FEL (red) and HSPC (cyan) architectures normalized by the maximum error in the naive system (feedback-control only). Robustness. (c) Mean angular position of FEL (red) and HSPC (cyan) during catch (N = 5 – solid) and regular perturbed trials (N = 5 – dashed). (d) Root mean square error (RMSE) in angular position during regular trained perturbed trials (FEL, HSPC, N = 5), catch trials (FEL-C, HSPC-C, N = 5) and extinction trials (FEL-E, HSPC-E, N = 10). Generalization. (e) Root mean square error (RMSE) in angular position during light-to-heavy generalization phases for FEL (red) and HSPC (cyan). ‘Light plant’ denotes the phase before plant perturbation. ‘Heavy plant’ denotes the phase after plant perturbation. (f) FEL mean angular position after plant perturbation (heavy plant – N = 10) without (dashed) and with the cue (solid) and after regular training with heavy plant (solid magenta). (g) HSPC: mean angular position after plant perturbation. As (f).