Motor Cortical Representation of Reaching and Grasping

Motor Learning in M1 and PMD

The motor system has an impressive ability to learn new skills and adapt movements to new environments. Mastery of a skill like playing the violin causes permanent changes in the brain areas controlling movement [1]. Yet, the motor system is also adept at learning skills on a much shorter time scale, such as learning to wield a new tool. This short-term motor learning process has been studied in humans using predictable force perturbations applied to the arm during reaching [2,3].

Over the course of ~100 reaches, research subjects learn to compensate for the perturbations to restore normal looking movements. Two common perturbations in these experiments are a "visuomotor rotation" (VR) [4], which offsets the visual feedback received about the movement, and the "curl field" (CF), a velocity-dependent force applied orthogonally to the hand movement direction (Figure 1).

We hypothesize that adaptation to these perturbations is mediated by altering the mapping between vision-centric planning coordinates and muscle-centric movement coordinates believed to occur in the motor cortices. We have previously shown that the relationship between the activity of neurons in primary motor cortex (M1) and the dynamics of movement remain fixed during curl field learning and thus cannot account for the adapted behavior [5]. Instead, learning may be mediated by altered recruitment of M1 neurons by upstream brain areas. We are testing this hypothesis by recording simultaneously from populations of neurons in M1 and dorsal premotor cortex (PMd). PMd plays a significant role in movement planning and is a strong source of inputs to M1. We compared the activity of neurons in these two areas as monkeys adapt to either VR or CF perturbations, and found that single cells in both M1 and PMd apparently maintain fixed relationships with movement covariates, either force or movement. In the curl field, for example, the "preferred direction" of M1 cells immediately rotate when the curl field is applied (Figure 2, blue) while PMd cells show no change at all (Figure 2, red). In neither area, is there a progressive change in tuning that can explain the slow behavioral adaptation. These results suggest that learning these perturbations does not result from changes within these networks. We are currently focusing on population-level approaches to understand how these brain areas work together to drive behavioral adaptation without changing the representations of single neurons. Through these experiments, we hope to shed light on the cortical mechanisms underlying motor learning, as well the functions of both M1 and PMd in coordinating movement.

↓ Jump to References

Figure 1. Monkeys perform a reaching task (top left) while we record from chronically implanted arrays in M1 and PMd (top right).The monkeys adapt to two kinds movement perturbations: the velocity-dependent curl field (bottom left) and the visuomotor rotation (bottom right)
Figure 2. Changes in "preferred direction" (PD) for single cells recorded from M1 and PMd as monkeys adapted to the curl field. We computed the change in PD relative to its direction in Baseline (before learning). Each dot represents the average change within the population of recorded neurons in a block of trials during (Adaptation) and after (Washout) learning. As behavioral adaptation occurs, we see unchanging representational tuning in both M1 and PMd throughout the learning process, with M1 tuning changing immediately with the onset of the altered dynamics, and PMd consistently representing the location of the goal.

Neural Representation of Uncertainty in Reach Planning

Figure 3.The effects of uncertainty on reach representations. A. A monkey made reaches, moving a handle in an effort to move a cursor from a center target to a target hidden on a surrounding ring. He was given visual cues consisting of series of radial lines, randomly drawn from two normal distributions (narrow and wide) centered on the target's location. Tightly clustered lines provided more information about the target location (low uncertainty) than did sparse lines (high uncertainty). Reach representations in PMd were noisier during high uncertainty trials than for low uncertainty trials, even if the movements were identical in direction and speed. B. Decoding the monkey's reach direction from PMd activity was much less successful for high uncertainty trials, as the activity did not "commit" to a single direction even as the movement was being made (left). However, decoding using M1 activity at the time of movement was unaffected by the uncertainty condition.

Every movement we make represents just one of many possible actions. It is an open question as to how the brain makes these choices, but it seems likely that the dorsal premotor cortex (PMd) is involved. When presented with two discrete reach targets, PMd appears to represent both options until one is shown to be the superior choice [6]. This has led some to speculate that motor decisions arise from a neural competition between movement plans within PMd [7-9]. However, it is not clear how well these interpretations extend beyond the multiple choice task.

To probe the role of PMd in situations without clear targets, we performed an experiment in which monkeys estimated the location of ambiguous targets represented by noisy visual cues. We found that the planning-related activity in PMd changed significantly with the noisiness of the cue, and that it appeared to reflect the monkey's subjective feeling of uncertainty about his decision [10]. Interestingly, these effects persisted even after a decision had been made (i.e., once a reach in a particular direction had started), an observation that appears to be different from the multiple choice condition. We are currently performing additional experiments to better understand the role of PMd in processing uncertainty and indecision during motor planning.

↓ Jump to References

Neural Population Activity Patterns During the Control of Movement

Figure 4. Neural manifolds for the control of movement. A. In the neural manifold view, the activity of each neuron results from underlying population-wide activity patterns, which we call neural modes. In this sketch, the differential contribution of each of the two latent variables to the neurons' activity is illustrated by the colors in each neuron's body. B. Spikes from three "recorded" neurons in the population. C. Latent variables underlying the generation of the neural activity. In this example, the activity of neuron N3 reflects latent variable L1, whereas neurons N1 and N2 result from weighted combinations of the two latent variables, L1 and L2 (see panel A). D. Neural population activity driving movement generation for two different tasks (black and purple trajectories). These neural trajectories mostly lie in a lower dimensional surface, in this case, two planes (shown in grey and violet). These low-dimensional surfaces are the neural manifolds that subsume neural population activity during movement generation for these two different tasks. Each neural manifold is defined by the population-wide activity patterns of the latent variables (shown as blue and green vectors for the grey manifold). Quite different neural trajectories (such as those shown in black and purple), can lie in manifolds that are remarkably close to each other. E. Principal angles between neural manifolds from two reach-to-grasp tasks (left plot, in black), and between all pairs of four different wrist tasks (right plot, colored traces). In both cases most of the manifold dimensions are significantly close, as they are smaller than the significance threshold (dashed grey line; P < 0.001). This result indicates that distinct behaviors are generated by remarkably similar neural population activity patterns.

Both humans and monkeys are able to perform a rich repertoire of movements by controlling the activation of the complex musculature of the arm and the hand. The movement planning problem is formidable, as the nervous system needs to generate the control signals to perform such diverse actions as finely manipulating a delicate object, or swinging a heavy tool. Primary motor cortex (M1) is one important area involved in limb motor control. For several decades, the neural control of movement has been studied one neuron at a time, using individual electrodes carefully placed in M1 while monkeys made repeated, simple movements [11]. While these studies have been informative, revealing much of what we know of the characteristics of the movement related signals in the brain, they can say very little about the detailed interactions within the networks of neurons that actually generate these signals.

To understand these mechanisms, we identified the "neural modes," the dominant activity patterns in the population of tens to hundreds of neurons we record from [12]. To this end, we used dimensionality reduction techniques (Cunningham & Yu, 2014), mathematical methods that identify the dominant patterns of covariance across neurons that explain most of the population activity (Fig. 4A-C). Remarkably, for our standard lab tasks, few neural modes (typically less than ten) explain most of the neural population activity [13].

Geometrically, those few neural modes define a manifold, a low-dimensional surface embedded in the full "neural space" whose axes are the activity of each individual neuron (Fig. 1D). We compared the geometry of the manifolds from different tasks (Fig 4D), and observed that they were very close to each other (Fig. 4E). This suggests that the neural activity driving the execution of different behaviors lives in a similar portion of the neural space. We also found that time-varying activation of those neural modes is quite similar across tasks, although there are also task-specific population neural modes.

We are currently trying to elucidate the relationship between these neural activity patterns and the commands to the muscles. Our first results indicate that the muscle activation patterns for different tasks are generated by combining task-specific and task-independent neural population activity patterns. We are also performing experiments that aim to map the neural manifold during free movement, while we record data continuously for many hours. Our ultimate goals are to understand: 1) the similarity of this neural manifold across tasks, and its stability over time, 2) how the neural manifold is sculpted by learning, and 3) how the activity within the manifold generates the muscle activation patterns that ultimately drive behavior.


[1] Elbert, T., Pantev, C., Wienbruch, C., Rockstroh, B., & Taub, E. (1995). Increased cortical representation of the fingers of the left hand in string players. Science, 270(5234), 305-7.
[2] Lackner, J. R., & Dizio, P. (1994). Rapid adaptation to Coriolis force perturbations of arm trajectory. Journal of Neurophysiology, 72(1), 299-313.
[3] Shadmehr, R., & Mussa-Ivaldi, F. A. (1994). Adaptive representation of dynamics during learning of a motor task. J Neurosci 14(5 Pt 2), 3208-24.
[4] Martin, T. a, Keating, J. G., Goodkin, H. P., Bastian, a J., & Thach, W. T. (1996). Throwing while looking through prisms. I. Focal olivocerebellar lesions impair adaptation. Brain, 119 ( Pt 4, 1183-98.
[5] Cherian, A., Fernandes, H. L., & Miller, L. E. (2013). Primary motor cortical discharge during force field adaptation reflects muscle-like dynamics. Journal of Neurophysiology, 110(3), 768-83.
[6] Cisek, P. & Kalaska, J. F. Neural correlates of reaching decisions in dorsal premotor cortex: Specification of multiple direction choices and final selection of action. Neuron 45, 801-814 (2005).
[7] Klaes, C., Westendorff, S., Chakrabarti, S. & Gail, A. Choosing Goals, Not Rules: Deciding among Rule-Based Action Plans. Neuron 70, 536-548 (2011).
[8] Coallier, E., Michelet, T. & Kalaska, J. F. Dorsal premotor cortex: neural correlates of reach target decisions based on a color-location matching rule and conflicting sensory evidence. J. Neurophysiol. jn.00166.2014 (2015). doi:10.1152/jn.00166.2014
[9] Pastor-Bernier, A. & Cisek, P. Neural Correlates of Biased Competition in Premotor Cortex. J. Neurosci. 31, 7083-7088 (2011).
[10] Dekleva, B. M., Ramkumar, P., Wanda, P. A., Kording, K. P. & Miller, L. E. Uncertainty leads to persistent effects on reach representations in dorsal premotor cortex. Elife 5, 1-24 (2016).
[11] E. V Evarts, "Relation of pyramidal tract activity to force exerted during voluntary movement.," J. Neurophysiol., vol. 31, no. 1, pp. 14-27, 1968.
[12] J. A. Gallego, M. G. Perich, L. E. Miller and S. A. Solla, "Neural manifolds for the control of movement," Neuron, under review.
[13] J. P. Cunningham and B. M. Yu, "Dimensionality reduction for large-scale neural recordings," Nat. Neurosci., vol. 17, no. 11, pp. 1500-1509, 2014.
[14] M. M. Churchland, J. P. Cunningham, M. T. Kaufman, S. I. Ryu, and K. V. Shenoy, "Cortical Preparatory Activity: Representation of Movement or First Cog in a Dynamical Machine?," Neuron, vol. 68, no. 3, pp. 387-400, 2010.
[15] P. T. Sadtler, K. M. Quick, M. D. Golub, S. M. Chase, S. I. Ryu, E. C. Tyler-Kabara, B. M. Yu, and A. P. Batista, "Neural constraints on learning.," Nature, vol. 512, no. 7515, pp. 423-6, Aug. 2014.