Interfacing with the computational brain

Abstract

Neuroscience is just beginning to understand the neural computations that underlie our remarkable capacity to learn new motor tasks. Studies of natural movements have emphasized the importance of concepts such as dimensionality reduction within hierarchical levels of redundancy, optimization of behavior in the presence of sensorimotor noise and internal models for predictive control. These concepts also provide a framework for understanding the improvements in performance seen in myoelectric-controlled interface and brain-machine interface paradigms. Recent experiments reveal how volitional activity in the motor system combines with sensory feedback to shape neural representations and drives adaptation of behavior. By elucidating these mechanisms, a new generation of intelligent interfaces can be designed to exploit neural plasticity and restore function after neurological injury.

Fig. 1

A. Schematic representation of multiple levels of redundancy within the motor system. B. Generating optimal movements requires advance knowledge of the environment. Here an inverse model converts visual target information into a feed-forward motor command to drive a BMI. C. Optimal feed-back control uses a forward model to generate predictions based on a copy of the motor command. This prediction can be combined with sensory feedback to form a state estimate that drives movement through a feedback controller.

Fig. 2

A. Learning curves for MCI performance. Average trial time reaches a comparable level irrespective of whether the decoding algorithm is intuitive (muscles act on the cursor in directions that are consistent with their action on the limb) or non-intuitive (muscles act in random directions). By the end of a single training session, subjects make fast, straight movements to the target. Adapted from [63]. B. Learning curves for BMI performance with the same neural population. Average trial time decreases over successive days for two monkeys. In this case the decoding algorithm was biomimetic (intuitive). Subsequent training on a randomized (non-intuitive) decoder required several more days to be optimized (not shown; see Fig. 6 in [13]). Adapted from [13]. C, D. Proposed model for learning MCI and BMI control involves acquiring the mapping between a copy of the efferent command signal and task-space feedback provided by the interface.

Fig. 3

A. Three (out of many) strategies that can be used to compensate for local rotational perturbation to a MCI or BMI mapping. The plots show the activity of representative units as a function of target direction relative to the original DoA (before a perturbation occurs). Tuning functions are assumed to be cosine-shaped, initially peaked at the DoA (gray line). Plots show predictions of three strategies (reaiming, reweighting and remapping) after a local perturbation in which the DoA of a subset of units rotates (indicated by arrows beneath the abscissa). Tuning functions for rotated units (solid line) and non-rotated units (dashed line) are shown following perturbation. B. Schematic of the hierarchical remapping framework. Remapping progresses from global to local levels of the redundant motor hierarchy. Therefore optimal adaptation to perturbations at a high level (e.g. visuomotor rotation) occurs before adaptation at lower levels (e.g. after local perturbation of BMI decoders).