24 Comments
Dec 29, 2023Liked by Sarah Constantin

Nice writeup! I’m gonna do the annoying thing where I self-promote my own cerebellum theory, and see how it compares to your discussion :) Feel free to ignore this.

My theory of the cerebellum is very simple. (Slightly more details here — https://www.lesswrong.com/posts/Y3bkJ59j4dciiLYyw/intro-to-brain-like-agi-safety-4-the-short-term-predictor#4_6__Short_term_predictor__example__1__The_cerebellum )

MY THEORY: The cerebellum has hundreds of thousands of input ports, with 1:1 correspondence to hundreds of thousands of corresponding output ports. Its goal is to emit a signal at each Output Port N a fraction of a second BEFORE it receives a signal at Input Port N. So the cerebellum is like a little time-travel box, able to reduce the latency of an arbitrary signal (by a fixed, fraction-of-a-second time-interval), at the expense of occasional errors. It works by the magic of supervised learning—it has a massive amount of information (context) about everything happening in the brain and body, and it searches for patterns in that context data that systematically indicate that the input is about to arrive in X milliseconds, and after learning such a pattern, it will fire the output accordingly.

Out of the hundreds of thousands of signals that enter the cerebellum time machine, some seem to be motor control signals destined for the periphery; reducing latency there is important because we need fast reactions to correct for motor errors before they spiral out of control. Others seem to be proprioceptive signals coming back from the periphery; reducing latency there is important for the same reason, and also because there’s a whole lot of latency in the first place (from signal propagation time). I’m a bit hazy on the others, but I think that some are "cognitive"—outputs related to attention-control, symbolic manipulation, and so on—and that reducing latency on those allows generally more complex thinking to happen in a given amount of time.

OK, now I’m going to go through your article and try to explain everything in terms of my theory. See how I do…

MOTOR SYMPTOMS: As above, without a cerebellum, you’re emitting motor commands but getting much slower feedback about their consequences, hence bad aim, overcorrection and so on.

ANATOMY: Briefly discussed at my link above, including links to the literature, but anyway, without getting into details, I claim that the configuration of Purkinje cells, climbing fibers, granule cells, etc. are plausibly compatible with my theory. It especially explains the remarkable uniformity of the cerebellar cortex, despite cerebellar involvement in many seemingly-different things (motor, cognition, emotions).

SIZE OF CEREBELLUM GROWING FASTER THAN OVERALL BRAIN SIZE IN HUMAN PREHISTORY: A big cerebellum presumably allows it to be a time-machine for more signals, or a better (less noisy) time-machine for the same number of signals, or both. Which is it? My low-confidence guess is "more signals"; I think it’s time-machining prefrontal cortex outputs (among others), and the number of such signals grew a lot in human prehistory, if memory serves. But it could be other things too.

CLASSICAL CONDITIONING: I don’t think your claim “the cerebellum is necessary and sufficient for learning conditioned responses” is true. I think it’s necessary and sufficient for eyeblink conditioning specifically, and some other things like that. For example, fear conditioning famously centers around the amygdala. I know you were quoting a source, but I think the source was poorly worded—I think it was specifically talking about eyeblink conditioning and "other discrete behavioural responses for example limb flexion", as opposed to ALL classical conditioning.

But anyway, for eyeblink conditioning, there seems to be a little specific brainstem circuit that goes (1) detect irritation on cornea, (2) use the cerebellum to time-travel backwards by a fraction of a second, (3) blink. Step 2 involves the cerebellum searching through all the context data (from all around the brain) for systematic hints that the cornea irritation is about to happen (a.k.a. supervised learning), and thus the cerebellum will notice the CS if there is one.

INDIVIDUAL PURKINJE CELLS CAN LEARN INFORMATION ABOUT THE TIMING OF STIMULI: If evolution is trying to make an organ that will reliably emit a signal 142 milliseconds before receiving a certain input signal, on the basis of a giant assortment of contextual information arriving at different times, then this kind of capability is obviously useful.

CEREBELLAR PATIENTS MAKE STRANGE ERRORS IN GENERATING LOGICAL SENTENCES: As above, I think there are cortex output channels that manipulate other parts of the cortex (via attention-control and such), and the cerebellum learns to speed up those signals like everything else, and this effectively allows for more complex thoughts, because there is only so much time for a "thought" to form (constrained by brain oscillations and other things), and the signals have to do whatever they do within that restricted time. I acknowledge that I’m being vague and unconvincing here.

ANTICIPATION: self-explanatory—this part is where you’re closest to my perspective, I think.

ELECTROLOCATION: I don’t know why electrolocation demands a particularly large cerebellum. Maybe latency is really problematic for some reason? Maybe the patterns are extremely complicated and thus require a bigger cerebellum to find them?

Expand full comment
author

this seems very much right.

do you have a source for the cerebellum being unnecessary for some kinds of classical conditioning (i.e. you can do fear learning with the cerebellum removed) or some other part of the brain being necessary as well for classical conditioning (i.e. classical conditioning fails when that other part is removed/lesioned)>

Expand full comment

Echolocation seems pretty latency-sensitive. Under water, you need a time resolution of about 1.3 milliseconds to perceive a distance resolution of 1 meter (the speed of sound in water is 1500 m/s -- something 1 meter further away will result in the sound needing to travel 2 additonal meters, which it can do in 1.33 milliseconds).

A cursory reading on electrolocation seems to indicate that the time resolution needed is between 100 microseconds and 1 millisecond, which is also very low latency.

Expand full comment

That’s interesting but I think unrelated to what I wrote above. I was talking about “sometimes it’s helpful for the latency to be very low” and you’re talking about “sometimes it’s helpful for the latency to be measured very accurately”. (Or sorry if I’m misunderstanding.)

Expand full comment

Perhaps relevant: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4830363/#sec1-7title

> Specifically, the participation of the cerebellum on the biological basis of time perception has been highlighted,13 but its function is yet not well established.83 It is believed that there are two systems of timing. The first, automatic, acts on motor circuits of the cerebellum is responsible by events of milliseconds.6,84 The second, controlled cognitively, is formed by parietal and prefrontal areas linked to attention and memory, being responsible by periods of minutes.85 A research analyzed patients with cerebellar lesion in tasks of discriminate time with intervals of 400 ms and 4s and noted damage into perception of milliseconds and seconds.28

Though, this does seem to imply there are some other functions than just time travel machine of the cerebellum, or that the time travel machine is used for other purposes, such as time perception.

Expand full comment
Dec 31, 2023Liked by Sarah Constantin

If you want some examples of the cerebellum in computational brain models, including modelling beyond just weights between neurons, the lab I did my master's in built some models https://github.com/ctn-waterloo/cogsci2020-cerebellum

The author of that paper would be happy to chat with you if you're interested, including how it features in a model of the motor control system https://royalsocietypublishing.org/doi/full/10.1098/rspb.2016.2134?rss=1

Expand full comment

So… the cerebellum is like a quick predictive-processing brain inside of your whole brain?

Expand full comment
author

Looks like it!

Expand full comment

VERY interesting. I learned a lot and got rid of some assumptions.

What might be any connection between cerebellar function and "motor" disorders, like "cerebral" palsy, Parkinson's, Huntington's, or even "Essential Tremor"? What about diabetes/excess blood glucose? Tourette's Syndrome? Restless Leg Syndrome?

H. Watkins Ellerson

Expand full comment

The cerebellum is the most ancient part of the brain involved in the most basic and necessary adaptations to the physical world. Great article.

Expand full comment

It explains benefit of proper true yoga.

Expand full comment

Have you investigated whether people on the autism spectrum have problems in their cerebellum? Some of the problems you describe seem to be problems my grandchildren on the spectrum have.

Expand full comment

This is a brilliant article. It sparks my thoughts regarding its ‘soft’ psychological implications. I wonder what portion of fast emotional reaction takes place in the cerebellum. Does it have implications for meditation, particularly Vipassana or Metta? Or in my field, intrusive thoughts and obsessional problems?

Expand full comment

Thank you for this insight.

My 8-year-old son has cerebellar hypoplasia, and what you’ve outlined here has been more helpful than anything we’ve heretofore been taught or exposed to in grasping the nature of his challenges.

Expand full comment

Great post. I want to comment first on one paragraph from your article:

"The only parts of the cerebral cortex without a corresponding cerebellar region are the auditory and visual cortices. (Suggestive, given that hearing and vision are passive senses, and my hypothesis in the previous section that the cerebellum has something to do with active sensing.)"

To state it explicitly, the three physical senses which do have corresponding regions are touch, smell, and taste. You refrained from saying these senses are truly active, and I agree with that sentiment. Academic papers claim the cerebellum affects eye motor control (https://www.frontiersin.org/articles/10.3389/fneur.2011.00053/full) and auditory processing (https://www.sciencedirect.com/science/article/pii/S1808869415307023), so the cerebellum affects all physical senses. Still, for some reason, a human being has no truly active senses.

Upon contemplation of this fact, a deeper point appears. It's true that humans don't generate a signal and measure the response with an active sensory organ. However, humans do have the ability to hold a belief (generate a hypothesis) and live according to that belief (test the hypothesis and draw conclusions). In exercising that ability, a normal human may employ all five physical senses. Crucially, a human may adjust their vantage point to attempt to receive information through at least one of the five physical senses.

Let's call that human ability "metasensing," where the physical senses may be moved to receive information according to a personal goal, which is a desire beyond survival and curiosity. Animals can obviously move around, and animals appear to have attentional capabilities. Animals will certainly act to survive, and a few species may exhibit what we may call curiosity. The difference of metasensing is a choice to seek information. Navigating to create a map and opening a book in pursuit of an answer are both examples of metasensing. Due to metasensation, all human senses may be used in an active manner. From where metasensation originates is not immediately obvious, but it certainly appears real.

If human senses may be used actively through metasensation, I disagree that the cerebellum is responsible for active sensing, but I do have a hypothesis for why vision and audio don't have a corresponding cerebellar region. As you state, while the cerebellum responds to stimulus and generates "predictions", the cerebral cortex (and perhaps other parts) will generate "perceptions". In my experience, auditory and visual information requires significantly more complex real-time processing structures and resources than taste, touch, and smell. The predominantly feedforward nature of the cerebellum is excellent for detection but poorly suited for recognition. My hypothesis is that while detection capabilities are handled in the feedforward-dominant cerebellum, audio and visual processing structures emerged in the cerebral cortex for recognition capabilities.

I also wish to comment on your last paragraph:

"The cerebellum may also inspire artificial-intelligence approaches somewhat, especially approaches to robotics or other control, in that it may be be beneficial to include a fast feedforward-only predictive modeling step to control real-time actions, alongside a slower training/updating pathway for model retraining."

Successful robotic control architectures typically act across multiple time scales, from microseconds to seconds to hours and more. For more information, I highly recommend reading the review article "Autonomous vehicle control systems — a review of decision making" (https://dx.doi.org/10.1177/2041304110394727). Two other interesting architectures you may want to investigate include the 4D/RCS Reference Model Architecture and the ASIMO Cognitive Map architecture.

Expand full comment

thankyou for this - it brought things flooding back to mind from my time in bristol (1976-7) working in a movement neurophysiology lab that focussed on spinal and cerebellar mechanisms in the control of movement and posture. My supervisor worked with decerebrate cats and was able to keep them walking with a near normal gait for up to 24 hours by stimulating the top of the spinal cord. So, the belief, at that time, was that the coordination of movement and gait was hard wired into the spinal cord, guided by input from stretch receptors and descending modulation, etc. Higher centres just said "walk" and the complex orchestration details of exact individual muscle tone were believed to be happening at a spinal level. My colleague and I were making microelectrode recordings of climbing fibre and mossy fibre cells in the cerebellum in response to peripheral electro stimulation. We saw the cerebellum as providing a homunculus model of action to compare intended with actual movement and adding in vestibular information, visual information and intention information. I always suspected that in some way it was involved in coordinating other information from higher centres. Your paper was an illuminating refresher - filling in that almost 50 year gap in my understanding! Thankyou.... Charlie Buck

Expand full comment

Does the Vestibular talk to Cerebellum directly?

Expand full comment

> Something needs to physically or chemically change in that neuron, representing the newly learned information, and causing the corresponding change in the neuron’s firing behavior.

Wondering, are there any hypotheses about how it does this?

Expand full comment
author

yes, lots, they even think they know which receptor is necessary to make it work. https://www.nature.com/articles/s41598-020-72581-8

Expand full comment

From a quick look at this as a non-biologist, I can't at all discern what the proposed/claimed medium/encoding is... any chance you'd be able to explicate that for us readers?

Expand full comment

Just scanning the article quickly, the basic gist is that the Purkinje cells involved in the eyeblink response fire non-stop (neurons have a few different firing patterns, either they fire in response to things, or they fire continually, or some mixture of the two). If they stop firing, the cerebellar nuclear cells fire and they generate a blink. Climbing fibers are projecting to the Purkinje cells, so if they fire, they deactivate the Purkinje cells. So our pattern of action here is Climbing fibers (inactive by default) -> Purkinje cells (active by default) -> cerebellar nuclei cells (inactive by default).

The mechanism by which these cells turn each other off or on involves usually involves firing a bunch of neurotransmitters into a synapse where they are then detected by different receptors, such as the GABA and AMPA-kainate receptors they mention. If agonists bind to these receptors, they change their status, like a gate, and allow (say) positive sodium ions to flood into the neuron, depolarising the neuron and creating an action potential. Or if the cell is always active, you might allow negative ions into the cell, or reverse the flow of negative ions out of the cell, in order to hyperpolarise it and deactivate the cell.

They are arguing that it is a specific protein complex (Gβγ) modifying a specific type of potassium channel (Kir3) that allows for this mechanism. I think the gist of the point is that these channels may allow for longer term modifications to that specific cell, as opposed to a typical neuron where activity can often just reset back to baseline.

I hope all of that helps! (Apologies for any errors in the above, it's quite late and I may have misread the article while skimming)

Expand full comment

Hm, I'm not sure it really answers my question, but helpful to know I guess? What I really want to know is like "where is the state stored"; what is it internally that actually changes over the long term.

Expand full comment

RE "bad news for anyone hoping to simulate a brain digitally", FYI I wrote a blog post recently making a similar point, using that same example (see my section 2.6) among others: https://www.lesswrong.com/posts/wByPb6syhxvqPCutu/examples-informing-my-pessimism-on-uploading-without-reverse

Expand full comment

I wonder what this means for Active Inference

Expand full comment