As an engineer, my instinctual concern is performance – does a device or technology work as expected? How effective is it? What are the performance bottlenecks and how do we improve it?
This last point requires a good understanding of the system in question, which often requires cycles of hypothesis testing to gather the knowledge of this system. In BMI, this is especially important. Understanding the underlying neural mechanisms of control and adaptation leads to drastic performance improvements in BMI. At the same time, BMI is also the best way to understand the underlying mechanisms.
This circular relationship is however quiet troublesome in designing BMI experiments:
- If we want to improve the performance of a wheelchair driving BMI, for example, we can come up with a new control scheme mapping the neural activities to the control space (wheelchair velocity and direction) and over sessions to see if the navigation time/distance decreases.
- If we want to see how a monkey adapt to a BMI decoder in driving a wheelchair, for example, we can come up with some control scheme and see how the monkey’s neural activities change as it becomes more proficient at the task
In the first scenario, we are implicitly assuming a certain underlying mechanism and designing the new control scheme based on that…purely performance oriented. However, do we attribute the resulting performance improvements on improved decoder design? It certainly seems forced when the common sample size of primate BMI experiements is 2 to 3. How do we account for different learning/adaptation style? Even if a new decoder design improves the performance for all monkey subjects, can we say that indeed it is a better design, regardless of the underlying learning mechanisms?
In the second scenario, how do we decide on what control scheme to use? Do we randomly assign weights to the recorded neurons? Derive the control scheme based on ? For decades we simply recorded neurons from the motor, premotor, and somatosensory cortices and assume that they will adapt. The performance improved based on the number of neurons recorded, and the type of filter thrown at the recordings.
Given assumption of different learning models, experiments designed to test the same decoder should probably differ as well. Control experiments where weights are randomly assigned to neurons might not be sufficient when assuming the intrinsic manifold model of learning.