The field of neuroprosthetics, or the control of prosthetic devices by brain activity, is a rapidly emerging one. It’s been previously shown that monkeys and humans can learn to control an on-screen cursor through the power of the mind alone, as their brain activity is either directly or indirectly measured and fed through a decoder that transforms the activity into cursor movement. One question we don’t yet know the answer to is: what is the nature of the neural activity that is decoded?
Ganguly et al. set out to answer this question. They trained two monkeys on a reach-to-targets task while recording the activity of a number of neurons. They found that a small group of these neurons were very stable across several days to weeks – that is, each individual neuron fired in the same way on day 19 as it did on day 1, for example. The researchers then used these stable neurons as inputs to their decoder. The figure below (Figure 2 in the paper) shows how the monkeys improved with time:
Improvement over time
It’s quite a complicated figure, but pretty straightforward to work out. Panel A shows the error rate (top) and time to reach the target (bottom) for one of the monkeys (the other is shown in the red inset). You can see that they get better and faster at hitting the target. Panel B shows that the monkeys get better as they make more reaches in a session on days 2 and 4, but by day 6 their performance is basically at a plateau. Panel C shows performance in the first 5 minutes of the task – note that after a few days their error rate drops significantly. Panel D shows the actual cursor movements they’re making, and how those movements get more correlated (i.e. more similar to one another) as time goes on.
The researchers also note that stable task performance is associated with the stabilization of the neural tuning properties – as the preferred directions of the neurons (the directions in which they fire the most) get more similar across days, performance increases.
So this is pretty interesting – a stable subset of neurons can be used to control an onscreen cursor. Where I think it gets cooler are some of the other questions they ask about this stable neural population. What happens if you remove certain neurons from the subset being fed into the decoder – that is, is the whole group being used to control the cursor, or just a subset of the subset? Here’s your answer (Figure 5 in the paper):
Performance relative to number of neurons dropped
Yup, performance drops significantly as you remove neurons. But, crucially, it isn’t destroyed with the loss of one or two, meaning that if a neuron dies, for example, it’s not going to radically change your performance. And since the brain is always learning and changing, other neurons are likely recruited to fill the gap.
There’s more in here, but one of the most interesting things is the ability to learn new decoders alongside old ones, and recall them rapidly when necessary, and also that the stable activity patterns emerge very early in each trial. This finding is very awesome because it parallels something I’m interested in for my own research: the idea of internal models in the brain that are used for certain tasks and can be switched between when necessary. So the same set of neurons can theoretically be used to perform different tasks, as long as the internal model ‘knows’ how to interpret their firing for each task.
As I said, incredibly cool stuff, and it means that we are another step closer to understanding how brain activity controls prosthetic limbs – and, of course, our real, natural limbs as well.
By the way, the paper is open access so you can read it yourself (if you don’t have an institutional subscription) via the link at the bottom.
Ganguly, K., & Carmena, J. (2009). Emergence of a Stable Cortical Map for Neuroprosthetic Control PLoS Biology, 7 (7) DOI: 10.1371/journal.pbio.1000153
Images copyright © 2009 Ganguly & Carmena