The muscles in your arms receive commands from the neurons in your brain. These commands are phrases of a complex language that neurons use to communicate with one another and, in the case of motor neurons, the muscles.. The neural language, like verbal languages, has groups of phrases that are thought to be redundant. This means that a brain commanding an arm to move could use various ‘neural phrases’ to do so. We exercise this flexibility often in our written language. For example, there are many friendly salutations that can end an email: sincerely…all the best…regards…peace. No matter what the rest of the email says, these phrases say the same thing to the reader, ‘good bye, this email is ending’. In theory there are practical implications for neural redundancy. A set of neural phrases that each lead to the same movement could be leveraged as a secondary code. In the context of the motor system, the patterns of activity could simultaneously be producing activity related to movement, while also doing another computation.
In their paper Constraints on neural redundancy, Jay Hennig and his colleagues wanted to understand the guiding principles for how neurons choose a specific phrase given so many equivalent options. One central challenge to asking these kinds of questions in neuroscience is that we are often able to only determine what a few neurons are doing at one time, and these few neurons are not the only neurons involved in producing a behavior. This means that knowing all the redundant phrases is impossible.
To address this, the authors use a powerful technique called a brain computer interface, or BCI. This BCI converts neural activity from a part of the brain that controls the arm into the movement of a computer cursor. Imagine if instead of moving a computer mouse with your hand, you could just think and your cursor would fly across the screen of your computer! The BCI can be used as a tool to study redundancy. Hennig uses the BCI to force an observed subset of neurons in the brain to use a language that is defined by researchers. The BCI is so important because under natural circumstances it is too hard for researchers to know the many possible redundant neural phrases before the experiment starts. One way to think about the BCI is as a new virtual arm with a motor-neural-language comprising vocabulary that the researchers defined.
To to build the BCI the authors placed tiny wires next to a few of the neurons in the motor cortex. Next, the authors defined some rules, the new language, to link neural activity with cursor velocity, e.g. if neuron A is active (see footnote) the cursor moves up, if neuron B is active the cursor moves right. All of these rules are enforced by the BCI and over time the brain learns to control neurons according to these rules so that the subject can make the computer cursor reach its intended goal. Hennig et al. designed the BCI to have built in redundancy. In effect the BCI takes all the redundant neural phrases and converts them to their essence (the intended cursor movement). The BCI will listen to many neurons in the motor cortex at once, making use of the patterns among them rather than just listening to one neuron at a time. For example, a BCI might define the speed of the cursor to be proportional to the activity of neuron C subtracted from the activity of neuron D. This kind of code allows for redundancy because there are many different ways to get a difference of two (e.g. 15 - 12 = 2 or 7 - 5 = 2). With this tool the research team can now ask how one redundant neural phrase is chosen over another.
The authors hypothesized that the activity of neurons could be guided by a few common-sense principles: for example, the brain should use as little energy as possible. Extending the example from above, we would hypothesize that neurons would have the lowest firing rates possible subject to their required difference. Surprisingly, the authors found these principles could not explain how the neurons chose their activity patterns. This means that during these experiments this brain area does not freely adjust its redundant activity. Instead, they discovered that although the BCI allowed for a broad set of redundant neural phrases, the redundant activity was yoked to the potent activity. In other words, they discovered that the best way to predict what was happening in the redundant dimensions was to look at activity in the potent ones.
While it is still possible that the brain could use the redundant dimensions for secondary processing, through their investigations Hennig et al. have constrained this theory. Their results suggest that the activity patterns that a neuron produces might be dictated by how they are connected. The connectivity of neurons could specify the patterns that one set of neurons can produce given what another set of neurons is saying. Before this work it was thought that redundancies could be freely exploited. Consider the email language example. Even though all email salutations say effectively the same thing, often times what is used depends on the other content of the message. If an email is addressed to a friend it could end with ‘see you later’ or ‘peace’, but if an email is addressed to an employer only ‘sincerely’ or ‘all the best’ or ‘regards’ are likely.
Footnote: Neurons send short pulses to each other. If a neuron is ‘active’ it is sending a lot of pulses per second (maybe 100), if it is ‘not-active’ it would send very few pulses per second (maybe 5).
Reference: Hennig, J. A. et al. (2018). Constraints on neural redundancy. eLife, 7:e36774