The general theme of the work presented in this dissertation has been to understand how learning occurs within the brain and more specifically how we can utilize learning to improve brain-computer interface performance. One of the main goals of BCI technology is to improve the quality of life for people who have limb-loss or people afflicted with paralysis or other neurological disorder. For a BCI to be truly effective, it should be easy for the person to use, and it should not require much intervention from a third party. To that end, it is important that we understand the neurological processes that underlie usage of a BCI. Are people able to use a BCI with an arbitrary mapping between neural activity and effector control? What are the processes that underlie learning in a BCI context? How can we design BCI mappings that best exploit the subject’s innate learning strategies? In order to develop BCIs for the clinic, much testing has been done and needs to continue to be done with animal subjects. It is important that we motivate the animals to demonstrate their best possible performance. This will help us to find weaknesses in the current implementation of BCIs. Can we design tasks that will motivate animals to demonstrate their best possible performance?
In Chapter 3, we tested the extent to which a subject could learn to use an arbitrary mapping between neural activity and cursor movement. We began each session by identifying the ‘intrinsic manifold’, which is the low-dimensional space within the high-dimensional neural space that captures the most prominent patterns of co-modulation among the recorded neurons (Figure 3.1). The monkeys then controlled the BCI cursor using an intuitive mapping that resided within the intrinsic manifold. In the middle of each session, we perturbed the mapping and observed whether the monkeys could learn to regain control of the cursor.
On some sessions, we perturbed the mapping within the manifold. That is, we reoriented the control space of the mapping so that it was different from the intuitive mapping but remained within the intrinsic manifold. On other sessions, we reoriented the control space of the mapping so that it was no longer within the intrinsic manifold. We found that the monkeys were better able to learn to regain control of the cursor on sessions when the control space was perturbed within the intrinsic manifold. This indicates that the monkeys could easily learn to recombine their natural neural activity patterns in new ways but that it was more difficult for them to learn to generate new co-modulation patterns among the neurons. In Chapter 4, we investigated the means by which the animals learned to control novel decoders. We found that the biggest changes in neural activity occur just after we switch from one BCI mapping to another. By projecting neural activity onto two planes of interest (the control plane and the null plane), we could decompose variability into its signal (i.e., variance across targets) and noise (i.e., variance within targets) components. Consistent with our intuition, as the signal increased in the control plane, the monkeys showed greater learning. Interestingly, the converse was also true. As signal increased in the null plane, learning decreased. In addition to analyzing neural activity as two-dimensional projections of the full neural space, we also analyze neural trajectories in the full neural space. We found that as the the trial-to-trial variability of the neural trajectories decreased, the monkey’s performance increased.
InChapter 5, we tested ways in which we could design decoders in order to elicit the best possible performance from animal subjects. Specifically, we incorporated the dimensionality reduction technique factor analysis into a Kalman filter. A standard Kalman filter maps neural activity to the effector kinematics (Wu et al., 2006). In our modified Kalman filter, we first reduced the dimensionality of neural activity to a set of latent factors, and then we mapped the latent factors directly to the kinematics of the cursor. We hypothesized that the dimensionality reduction step would extract latent dimensions that captured underlying high-level control signals in the neural activity that gave rise to the recorded spike counts. We were interested in finding the number of latent dimensions that would lead to the best performance. As we increased the number of latent dimensions in the FA model, we found that the monkey’s performance also increased. However, performance with the standard
Kalman filter was not substantially different from the peak performance with the modified Kalman filter. This indicates that if there are low-dimensional control signals present in the neural activity, the standard Kalman filter does a sufficient job of identifying those to move the cursor.
Finally, in Chapter 6, we explored whether we could motivate monkeys to demonstrate better performance with a BCI by changing the requirements of the task. We used the ‘instructed path task’ for the first time in a BCI setting. In the instructed path task, the monkeys were required to move the cursor along a path that was displayed to them. For both monkeys, we tested their performance on straight instructed paths and single-inflection paths, while one of the monkeys was able to control a cursor on even more complicated paths. To enable comparison, we also trained both monkeys on a point-to-point task that was akin to center-out. We compared each monkey’s performance on the point-to-point task to his performance on the straight instructed path task. We found that the presence of the instructed path motivated one of the monkeys to demonstrate better performance relative to the point-to-point task. The other monkey demonstrated the same level of performance in both tasks. From this, we concluded that there is a need to tailor BCIs to be specific to each individual.