Lots of what neurons do adds up to zero: how the cerebellum controls muscle movement.
Our brains are always active. Meditating, relaxing, sleeping ... whatever we are up to, the brain doesn't take a holiday, except in the deepest levels of sleep, when slow waves help the brain to reset and heal itself. Otherwise, it is always going, with only slight variations based on computational needs and transient coalitions, which are extremely difficult to pick out of the background (fMRI signals are typically under 3% of the noise). That leads to the general principle that action in a productive sense is merely a tuned overlay on top of a baseline of constant activity, through most of the brain.
A recent paper discussed how this property manifests in the cerebellum, the "little" brain attached to our big brain, which is a fine-tuning engine especially for motion control, but also many other functions. The cerebellum has relatively simple and massively parallelized circuitry, a bit like a GPU to the neocortex's CPU. It gets inputs from the places like the the spinal cord and sensory areas of the brain, and relays a tuned signal out to, in the case of motor control, the premotor cortex. The current authors focused on the control of eye movement ("saccades") which is a well characterized system and experimentally tractable, in marmoset monkeys.
After poking a massive array of electrodes into their poor monkey's brains, they recorded from hundreds of cells, including all the relevant neuron types (of which there are only about six). They found that inside the cerebellum, and inside the region they already knew is devoted to eye movement, neurons form small-world groups that interact closely with each other, revealing a new level of organization for this organ.
More significantly, they were able to figure out the central tendency or vector for Purkinje (P) cells they ran across. These are the core cells of the cerebellar circuit, so their firing should correlate in principle with the eye movements that they were simultaneously tracking in these monkeys. So one cell might be an upward directing cell, while another one might be leftward, and so forth. They also found that P cells come in two types- those (bursters) that fire positively around the preferred direction, and others (pausers) that fire all the other times, but fire less when eyes are heading in the "preferred" direction. These properties are already telling us that the neural system does not much care to save energy. Firing most of the time, but then pausing when some preferred direction is hit is perfectly OK.
 |
| Two representative cells are recorded, around the cardinal/radial directions. The first of each pair is recorded from 90 degrees vs its preferred direction, and adding up the two perpendicular vectors (bottom) gives zero net activity. The second of each pair is recorded from the preferred direction (or potent axis), vs 180 degrees away, and when these are subtracted, a net signal is visible (difference). Theta is the direction the eyes are heading. |
Their main finding, though, was that there is a lot of vector arithmetic going on, in which cells fire all over their preferred fields, and only when you carefully net their activity over the radial directions can you discern a preferred direction. The figure above shows a couple of cells, firing away no matter which direction the eyes are going. But if you subtract the forward direction from the backward direction, a small net signal is left over, which these authors claim is the real signal from the cerebellum, which signals a change in direction. When this behavior is summed across a large population, (below), the bursters and pausers cancel out, as do the signals going in stray directions. All you are left with is a relatively clean signal centered on the direction being instructed to the eye muscles. Wasteful? Yes. Effective? Apparently.
 |
| The same analysis as above, but on a population basis. Now, in net terms, after all the canceled activity is eliminated, and adding up the pauser and burster activity, strong signals arise from the cerebellum as a whole, in this anatomical region, directing eye saccades |
Why does the system work this way? One idea is that, like for the rest of the brain, default activity is the norm, and learning is, perforce, a matter of tuning ongoing activity, not of turning on cells that are otherwise off. The learning (or error) signal is part of the standard cerebellum circuitry- a signal so strong that it temporarily shuts off the normal chatter of the P cell output and retunes it slightly, rendering the subsequent changes seen in the net vector calculation.
A second issue raised by these authors is the nature of inputs to these calculations. In order to know whether and how to change the direction of eye movement, these cells must be getting information about the visual scene and also about the current state of the eyes and direction of movement.
"The burst-pause pattern in the P cells implied a computation associated with predicting when to stop the saccade. For this to be possible, this region of the cerebellum should receive two kinds of information—a copy of the motor commands in muscle coordinates and a copy of the goal location in visual coordinates."
These inputs come from two types of mossy fiber neurons, which are the single inputs to the typical cerebellar circuit. One "state" type encodes the state of the motor system and current position of the eye. The other "goal" type encodes, only in case of a reward, where the eye "wants" to move, based on other cortical computations, such as the attention system. The two inputs go through separate individual cerebellar circuits, and then get added up on a population basis. When movement is unmotivated, the goal inputs are silent, and resulting cerebellar-directed saccades are more sporadic, scanning the scene haphazardly.
The end result is that we are delving ever more deeply into the details of mental computation. This paper trumpets itself as having revealed "vector calculus" in the cerebellum. However, to me it looks much more like arithmetic than calculus, and nor is the overall finding of small signals among a welter of noise and default activity novel either. All the same, more detail, and deeper technical means, and greater understanding of exactly how all these billions of dumb cells somehow add up to smart activity continues to be a great, and fascinating, quest.
- How to have Christmas without the religion.
- Why is bank regulation run by banks? "Member banks ... elect six of the nine members of each Federal Reserve Banks' boards of directors." And the other three typically represent regional businesses as well- not citizens or consumers.
