It is true, I never hear talk about the cerebellum here at Numenta. Let us know if there are any interesting findings about how feedback from the cerebellum connects into neocortex.
from what i’ve read here this could even be dubbed as the 7th hidden layer of the neocortex
The structure of the cerebellum includes several layers of varying sized neuron groupings such as would easily be applied to time delay circuits or time delay channels.
I think of it as a non-specific multitimer resource or SenseOfTiming organ that can be used by the associative learning process to blend in our natural sense of timing with our experiences of sequenced movements.
Learning muscle group activation in concert with duration is the essence of coordination, but also the whole spectrum of dance and rhythm is serviced by this in-skull timing sensor/peripheral organ.
I note that the book, On Intelligence, downplays the significance of this organ (cerebellum), and within Numenta, there is a strong tendency to follow the eureka moment that points to sequence memory as a cortical layer issue. This eureka may be premature, and may miss the big picture.
Two aspects are important in sequence memory:
- the association of what happens together (the global cortical engram of a moment of experience) becomes linked to what was experienced immediately before - i.e. the trailing signals of a previous moment are still active enough to be part of the fixation of the next momentary engram.
- the discrete sense of continuous time changing between one engram and a related one - which can reasonably be from input from the chorus of timing sense that is generated in the cerebellum.
Some memory objects seem very distinct or not associated with any particular timing’s so for these tapping into the cerebellum or establishing a timing feedback circuit may not get in as part of the engram, but nearly everything with linguistic memory or verbal cogitation would have taps into the multitimer organ, or, in the case of deficient cerebellum, plasticity can allow another part of the brain to generate timing sensibility for associative memory formation and recall.
The cerebellum in widely acknowledged in it’s role in coordinating motion.
Less well understood is some of the other roles in cognition and language.
The Cerebellum is primarily connected to the motor cortex. Most think of this for the actuation of the body.
I should mention that it also modulates the connections from the motor and pre-motor projections to various other sections of the brain and allow the motor neurons to act as as drivers for action in the brain, prompting new states that are eventually perceived by the the temporal lobe as what we usually call “thoughts.”
The cerebellum facilitates the generation of longer and more stereotyped (Smoother?) sequences.
Some interesting papers on cortical-cerebellum connections:
I find this very interesting: From there, projections are sent to cerebral structures e.g. motor cortex. In turn, the cerebral cortex sends projections to pre-cerebellar nuclei, thereby forming multiple ‘cortico-cerebellar’ loops, reminiscent of cortico-basal ganglia loops
CEREBELLUM, BASAL GANGLIA AND CORTICAL CONNECTIONS UNMASKED IN HEALTH AND DISORDER
An important function of the cerebellum is controlling eye movements. A symptom of cerebellum damage is nystagmus.
This extract of an article from Christof Koch may interest you.
Key points :
- One can live a nearly-normal life without a cerebellum
- The circuitry of the cerebellum is almost exclusively feed-forward (no complex feedback loops)
- 2d-connectivity of Purkinje neurons in the cerebellum, versus 3d-connectivity in the neocortex
Koch uses this lower degree of complexity of the cerebellum to justify its uselessness role in consciousness.
Christof Koch - May 2018 : https://www.nature.com/articles/d41586-018-05097-x
Or consider the cerebellum, the “little brain” underneath the back of the brain. One of the most ancient brain circuits in evolutionary terms, it is involved in motor control, posture and gait and in the fluid execution of complex sequences of motor movements. Playing the piano, typing, ice dancing or climbing a rock wall—all these activities involve the cerebellum. It has the brain’s most glorious neurons, called Purkinje cells, which possess tendrils that spread like a sea fan coral and harbor complex electrical dynamics. It also has by far the most neurons, about 69 billion (most of which are the star-shaped cerebellar granule cells), four times more than in the rest of the brain combined.
What happens to consciousness if parts of the cerebellum are lost to a stroke or to the surgeon’s knife? Very little! Cerebellar patients complain of several deficits, such as the loss of fluidity of piano playing or keyboard typing but never of losing any aspect of their consciousness. They hear, see and feel fine, retain a sense of self, recall past events and continue to project themselves into the future. Even being born without a cerebellum does not appreciably affect the conscious experience of the individual.
All of the vast cerebellar apparatus is irrelevant to subjective experience. Why? Important hints can be found within its circuitry, which is exceedingly uniform and parallel (just as batteries may be connected in parallel). The cerebellum is almost exclusively a feed-forward circuit: one set of neurons feeds the next, which in turn influences a third set. There are no complex feedback loops that reverberate with electrical activity passing back and forth. (Given the time needed for a conscious perception to develop, most theoreticians infer that it must involve feedback loops within the brain’s cavernous circuitry.) Moreover, the cerebellum is functionally divided into hundreds or more independent computational modules. Each one operates in parallel, with distinct, nonoverlapping inputs and output, controlling movements of different motor or cognitive systems. They scarcely interact—another feature held indispensable for consciousness.
Yes, people can live without one.
You can learn a lot about that role by watching someone who’s been pulled over for drunken driving, Schmahmann says. “The state trooper test is a test of cerebellar function. So the effect of alcohol on cerebellar function is identified by everybody who’s ever done walking a straight line or touching their finger to the nose.”
As far as the common trope that you live a mostly normal life:
I guess it depends on what you consider normal. I would think that going though life as if I was perpetually very drunk would be a drag. I suppose you could get used to it; actual drunks do.
Thanks Bitking for the articles.
I think this extract from your first link is a good summary of the cerebellum utility:
Research on Jonathan and people like him supports the idea that the cerebellum really has just one job: It takes clumsy actions or functions and makes them more refined. “It doesn’t make things. It makes things better,” Schmahmann says.
“What we now understand is what that cerebellum is doing to movement, it’s also doing to intellect and personality and emotional processing,” Schmahmann says.
are there clues out there on how this all ties into HTM theory?
The lines going out and back to the cerebellum are (respectively) output from the field of HTM activated columns, and the return to the HTM inputs as feedback to improve the on-going sequential response.
Sounds like an optimization.
Yes - the difference between drunken out-of-control you and finely tuned Olympic athlete you.
That could be described as optimized.
Cerebellum adds detail to to whatever frontal cortex is doing. Cortex is a big-picture part: it’s networking is relatively sparse but long-range. Cerebellum is the opposite: lots of neurons packed in a small space, so their connections are short. It worked on it’s own in lower animals, which live in here-and-now.
In mammals, cortex is the boss that outsources low-level work to cerebellum.
I have some sources and speculations on similar density vs. range trade-offs within the cortex:
http://cognitive-focus.blogspot.com/ , top post.
These papers are a very good starting point in understanding the function of the cerebellum. There has been considerable additional work since then but the original work has basically stood up very well.
A theory of cerebellar cortex - David Marr
David Marr’s theory of cerebellar learning: 40 years later - Piergiorgio Strata
A theory of cerebellar function - James S.Albus
MarrAlbus Model of Cerebellum - Computational Models of Neural Systems
Lecture 2.2 - David S. Touretzky - September, 2013
Not quite. That’s a good description of the granule cells in the cerebellum, but completely ignores the Cerebellar Purkinje cells, which have by far the largest dendritic trees of any neuron. They can often have as many as a million synapses.
It seems to me that the granule cells are doing some kind of dimensionality increase on the input data to make it sparser and pick out simple patterns, and then feed that data into the Purkinje cells, which proceed to learn patterns across extremely large amounts of data.
It seems to me that the Cerebellum is there to integrate data across very large portions of the cortex - perhaps as an optimization to the horizontal connections in the Thousand Brains Model? Being able to recognize complex, large-scale patterns across cortical data seems pretty useful.
One other thing to note is that Purkinje cells are known to perform supervised learning; there are special axons that connect to Purkinje cells called Climbing Fibers. They seem to be driven by the cortex and brain stem, and directly cause large, complex spikes in Purkinje cells.
You probably know a lot more about Cerebellum, but just going by raw numbers, it seems to have > 10x neurons per volume compared to cortex. That density must come at some cost, and main cost here are connections and their metabolic activity. Not simply the number of connections, but their average length and activity. So, Purkinje cells have huge dendritic trees, but I guess their axons are a lot shorter or less active. Correct me if that’s wrong, but there must be some trade-off.
Regarding functionality, I know that Cerebellum can only detect short-term dependencies, < 4 seconds?
I am guessing that spatial correlations that it can detect are also short-range, relative to cortex.
This goes back to my point about generalization (range) vs. detail (density) trade-offs.
The trade-off seems to be that there just aren’t very many of them. There are ~16 billion pyramidal neurons in the cortex. There are ~50 billion granule neurons in the cerebellum. There are only ~10-15 million Purkinje neurons in the cerebellum.
There doesn’t seem to be any feedback from the Purkinje cells (Cerebellum output) back to the granule cells or Purkinje cells. If the cerebellum is doing any sequence learning, it’s quite different from how the cortex does it. In fact, based on the circuits in the cerebellum, I think the Cerebellum likely doesn’t do any sequence learning, at least not on its own.
My guess as to what’s going on is this: the cerebellum is basically like a very long-range version of a spatial pooler with some level of support for reinforcement learning. Axons from the cortex and brain stem project to the granule cells, which don’t do any learning, but just increase sparsity. Their axons (called parallel fibers) project to the dendrites on Purkinje cells, which can be thought of as working like pyramidal neurons with only proximal dendrites. If any sequence learning is going on there, it’s likely just from the Purkinje cells deriving sequence-related patterns from the input from the cortex.
The resulting outputs from the Purkinje cells then project back to the brain stem and cortex (indirectly, they actually project to some nuclei in the brain stem, which then project to other places). The cortex and brain stem may also send feedback to the climbing fibers which directly trigger Purkinje cells, allowing them to train them for specific responses.
I can agree there I don’t see anything in the cerebellum to create a sequence.
I do have to point out that it is being driven by cortex that is known to have waves of activation.
If the lines going into the cerebellum are dynamic there is no need for a sequence generation in the cerebellum itself.
The Cerebellum - UBC Neuroanatomy
In the cerebellum neurons are in a large number parallel GROOVES
while in the neocortex neurons are in a folded sheet.
In the cerebellum neurons make few connections (100+) along a single direction
while in the neocortex neurons make many connections (1000+). in any direction.
In the cerebellum adjacent neurons are related to actions or events happening chronological after each other (close to each other in time)
while in the neocortex adjacent neurons are related to actions or events that are similar (close to each other in space) e.g. “pixels” on the retina.
In the cerebellum neurons are programmed by input from the neocortex (motor output of neocortex is input for cerebellum)
while in the neocortex neurons are programmed by input from the senses (perception like vision, touch, hearing, taste, balance,…)
In the cerebellum neurons are mainly used for procedural and time sequential execution
while in the neocortex neurons are mainly used for recognition and reasoning (association and correlation and simulation)
In the cerebellum neurons are arranged in a topology based on time
while in the neocortex neurons are arranged in a topology based on space.
In the cerebellum perception input is used to activate responses (such a response can be starting a sequential thread in the cerebellum)
while in the neocortex perception input is used to synchronize sequential steps (each next step is waiting for feedback from touch or muscle position).
In the cerebellum neurons pulse source input (from the hippocampus) is used to enforce a specific type of recognition (e.g. searching visually for your car-keys feeds the visual car key qualia neuron group)
while in the neocortex pulse source input (from the hippocampus) is used to activate a single step in sequential thread in a neuron groove (pulse stream is directed to the next step by neurons that fire on feedback)
NOTE: do not try google this and come complaining when you don’t find it confirmed (Digitronic proprietary research results have never been published). This description is our personal vision and therefor I do not claim nor request anybody to agree.
(Mark A Peaty) Stephan, I have not yet watched the video [I have to go out in a minute] but just want to say I read an article some time ago which seemed to make very good sense to me.
It said that the cerebellum acts as a monitor of outcomes resulting from cortical command outputs. The crux of the idea is that if the resulting movement, perception, whatever aligns with expectations, ie is the same as previous results of those commands, then the feedback from muscles, etc is prevented from reaching the cortex thus freeing up cortex to deal with unexpected outcomes.
This makes a lot of sense to me; in effect the cerebellum is seen to be guardian of useful habits.
(Stephan Verbeeck) Mark A Peaty That is about right, but not 100% correct.
The darwinistic advantage of having a cerebellum is that it indeed offloads the neocortex from tedious motor control actions allowing you to concentrate on other things than lifting foots alternating left right while walking. Without the cerebellum you would not be able to walk properly without focusing on walking (which leaves you pretty defenseless because that would require doing complicated upper body motor commands at the same time).
So the cerebellum learns from the sequence of actions that you do deliberately and builds neuron pathways that can do the same actions in the same sequence but then without the neocortex doing more than here and there gently giving a nodge to slightly correct the cerebellum sequential thread of execution.
But as to quality correction it is the other way around.
It is the neocortex that has to intervene and correct when the sequence (e.g. dancing) was not correct so if you want to alter a previously learned motor sequence then again you have to focus on it and deliberately contradict (overrule) the cerebellum actions and the cerebellum will then also learn these correction (integrate them in the existing sequence/thread of neurons firing).
this is excellent information, IMO especially because of your #1 observation,
“In the cerebellum neurons are in a large number parallel GROOVES while in the neocortex neurons are in a folded sheet.”,
I say (my theory) that what the cerebellum does is to behave as
- a multitimer peripheral sense organ to the neocortex (and underlying structures) such that durations (from fractions of a millisecond to multiples of days) can be sensed via the cerebellum as it marches through it’s ordinary function, [when any engram forms, it carries a start time attribute into the cerebellum, and when the engram is significantly succeeded by another, the other engram is provided with an ‘offset’ timing attribute that is the difference between ‘now’ and the ‘start’]
- a delay circuit which can be complex-ed with motor behavior to facilitate coordination, by adding appropriate waits to the next step in practiced sequenced movements.
timing requires bland linear circuitry such as the parallel grooves enables.