Intracortical neural interface systems have demonstrated that a tetraplegic human was able to obtain continuous two-dimensional control of a computer cursor using neural activity recorded from his motor cortex. In this paper from 2008, Neural control of computer cursor velocity by decoding motor cortical spiking activity in humans with tetraplegia, some design choices to improve such a system were evaluated, including the kinematic representation for cursor movement, the decoding method that translates neuronal ensemble spiking activity into a control signal, and the cursor control task used during training for optimizing the parameters of the decoding method.
(Note: Tetraplegia is paralysis caused by illness or injury to a human that results in the partial or total loss of use of all their limbs and torso).
Basically, the researchers implanted electrodes in the patients’ motor cortex which can monitor the firing rates of the neural units (groups of neurons). In training, the subjects were either instructed to imagine moving a cursor to a random point on the screen, or imagine moving cursor from the center to one of the four cardinal targets on the screen. In this “open-looped” training, the subjects simply try to duplicate the movements of the cursor on screen in their mind. During the “closed-loop” tests, the subjects’ imagined movements are shown on the screen as a result of what the decoding algorithms interpret their neural firing rates.
The two different tasks were meant to determine what’s the optimal kinematic representation for cursor movement. In the first task, the patients were told to imitate the cursor’s position; the second task instructed the patients to imitate the cursor’s velocity, all in their minds. In other words, the first task tested if decoding the neural firing rates based on position yield more accurate results; the second task tested velocity.
Two decoding algorithms were used, a linear filter (for position and velocity), and a Kalman filter (for velocity only).
There were many interesting results, but the one that stood out to me the most was that:
From the statistics of the CC (correlation coefficients) measures, we found that the firing rates of the majority of units in both participants were more strongly correlated with cursor velocity than with position. The statistical test showed that 65.3% of units were significantly correlated (t-test;
p < 0.01) with both position and velocity, 5.7% with position
alone and 24.7% with velocity alone. This result indicated
that 95.7% of all the units detected by the microelectrode
array in human motor cortex were significantly correlated
with at least one of the TC kinematic parameters. The CC
values for velocity were overall larger than those for position
(p=0.01; Wilcoxon rank sum test).
In 16 of 17 recording sessions, more units were significantly
correlated with cursor velocity than position. Furthermore, among all 1226 units, 69.9% of units exhibited stronger correlation with velocity while only 25.9% of units showed stronger correlation with position. This stronger correlation with velocity was observed regardless of the training task hese findings suggest that, for these neuronal populations, a computer cursor may be better controlled by a person with tetraplegia if using a decoding method based on cursor velocity rather than cursor position.
These findings suggest that, for these neuronal populations, a computer cursor may be better controlled by a person with tetraplegia if using a decoding method based on cursor velocity rather than cursor position.
So it seems that neural firing rates are correlated with velocity more than with position, at least as far as the tests used are concerned. But apparently the superiority of velocity decoding over position decoding has been observed in previous non-human primate studies as well:
In particular, the work of Taylor et al and following work on real-time closed-loop cortical control of a prosthetic device (e.g. a robotic arm) by neurologically intact non-human primates
established the feasibility of using velocity decoding from
neural population activity to control prosthetic devices. Our
finding that decoding velocity provided better cursor control
than decoding position supports these previous findings and
extends them to humans. Beyond previous work, the present
study also shows that cursor velocity was better extracted
from the motor cortical neurons during purely imagined (or
attempted) movements in tetraplegic humans. This was
evaluated with an explicit comparison between velocity and
position decoding during closed-loop cursor control.
Again we see the stronger correlation between cortical neuron firings with velocity. Let’s refer back to my first Muscle Memory post, where I mentioned that the motor cortex stores representation of a movement in terms of either velocity or position, then either integrates or derives to complete information needed for the central-nervous-system (CNS)’ proportional-plus-derivative (PD) controller.
This study in neural control seems to suggest that the neurons might indeed store the velocity representation of a movement, manifested in higher correlation between velocity and firing rates, as well as smoother control (although that can be attributed to the actual filtering algorithms used). Further, as quoted above, reasonable cursor control was achieved more easily with velocity-based algorithm, suggesting the native representation of movements in the motor cortex (again the algorithm can be confounding).
So to recap with my emerging, naive understanding of muscle “memory” (a misnomer…should be more precisely be called learned movement proficiency):
When we first learn a movement, we try to imitate the positions, as the velocity is harder to discern and vary more across individuals doing the same movements. As we get more proficient in duplicating the positions, the motor cortex “remembers”, or become tuned to the velocity signals generated. In further executions of the movements, these velocity signals become reference for the CNS PD controller.
When we start learning new movements, most of us observe the position of someone else doing the movement, and in turn the instructor usually communicates the movement in terms of position. As we start to become familiar with the motion, we start thinking of the movement’s dynamics–the amount of force and speed necessary, instead of its kinematics–the distance and angle that define a movement in space.
I suppose this is empirical evidence for my version (I shall check against more established theories!) of “muscle memory”…imitation of the kinematics comes first in the learning process because of the ease to observe and communicate them; dynamics learning then takes over because that’s the native representation of movement in our brains.
The big idea, I guess, is learning a movement’s velocity profile might yield faster results.
I think several things can be learned through these results:
1) Supplementing movement instruction with velocity information may improve learning rate. This doesn’t necessarily apply to movements where velocity control is not standardized (e.g. people might do six-steps differently for style effects). Rather, velocity description is more helpful in learning movements that have established “forms” or execution, such as a volleyball spike, basketball shot, or swim strokes.
2) The previous point might seem obvious. Perhaps more importantly, the results of these researches provide a strong basis for further work on brain-machine interface (or Ironman suit). Knowing the native/dominant representation of movements in the brain allows for effective signal decoding. A more sophisticated decoding algorithm based on position, velocity, and acceleration (maybe even jerk!) can probably be made, giving heavier weight to the velocity information, to yield more accurate neural translations.