Just this morning, it was such a great pleasure when two of the folks at Leading Edge Forum discussion asked me "By the way, what DO we know about how brains work that 99% of the world [especially in computer industry] does not know and needs to know?" I wish I had asked them to record it, because it was more direct and human than anything I have seen really addressing this.
"The first thing they need to know is the principle of MASS ACTION. They show you these little pictures of "how the brain works and what is there" which are basically just colorful maps of the outer surface of the neocortex, telling you what the function is of each colorful region. What they often miss is that: (1) the neocortex, that top outer part of the brain, is only about HALF of the brain by weight; (2) those colorful diagrams what you see in adult brains after learning, and do not account for the fact that ANY part of the neocortex can learn the usual roles of any other part, if the relevant inputs are available and there is an impetus to learn.
"Maybe the greatest breakthrough ever in systems neuroscience was the work of Karl Lashley, demonstrating this principle of mass action, demonstrating that the neocortex is a kind of UNIVERSAL LEARNING machine, not just a collection of ad hoc apps thrown on the floor. Two of his students, Karl Pribram and Walter Freeman, were the world's main leaders in systems neouroscience, in my view, until they died of old age. I am grateful that I was able to work closely with both. (Robert more with Freeman, me more with Pribram, but we all connected.)
Next they asked: Is the brain like a neural network at all? Does it do the prediction and pattern recognition stuff?
"Probably the two CURRENT leaders of systems neuroscience are Miguel Nicolelis and Buzsaki. I met Nicolelis in one of Pribram's
workshops, before he became famous, when he showed us that LEARNING TO PREDICT is one of the core universal learning functions which guides the cortex to learn many of the more specific processing tasks you see in those color diagrams.
MORE PRECISELY: there is a fairly direct hard-wired path from our retinas to the main part of a small organ called the thalamus, in the very center of the brain, a kind of central buffer or switchboard. We learned long ago about the cells which act as a kind of "movie screen of the brain", showing a new image every tenth or eight of a second or so. We learned that there are many fibers carrying that image to the neocortex, which is studying and analyzing that movie using a warehouse of tools much larger than the movie itself.
What Nicolelis showed is that there is a dense network of fibers coming BACK from neocortex to thalamus, supporting ANOTHER set of cells which LEARN TO PREDICT the raw input cells. The first universal function of neocortex is to PREDICT these current raw inputs as a function of earlier information. that is the number one universal learning principle/function, which causes the more detailed stuff to emerge by learning.
But that is not the ONLY universal learning function of neocortex. Think of neocortex as something like a company which makes two or three products, and maximizes its total revenue from all two or three. It is like the UNIVERSITY OF THE BRAIN, the main site of deep analysis, modeling and prediction, but it also produces a kind of CREATIVE IDEAS, OPTIONS for actions.
Deep under the neocortex are many important things, including a blobby gooey set of neural circuits called "basal ganglia."
When I first took neuroscience at Harvard (1964), people knew they had something to do with muscle control, but not a lot.
Now we know they are like the CORPORATIONS of the upper brain, organizations which take value inputs form many customers or stakeholders, and take or order big actions like movement plans and skills. The decision of WHAT TO DO is mainly made there, but they depend on the OPTIONS, POSSIBLE actions, suggested to them by the "universities," the neocortex.
{As I type this... some corporations just follow the suggestions, while others really evaluate... it depends on what they have learned about what works, but they decide.)
Then they asked: "But where are time and space in this picture?"
In truth, I feel I know the answer to this more than anyone else on earth, because of what we published in Werbos and Davis,
based on analysis of the Buzsaki data,
BUZSAKI is best known for his work on a DIFFERENT part of the brain, not neocortex or thalamus or basal ganglia. He is best known as leader in understanding the HIPPOCAMPUS, another key part of the higher brain., Like the basal ganglia, it is better known now than it was on 1964, and Buzsaki gets a lot of the credit. A cynical view is that Lashley, Pribram, Freeman and others had already explained so much of the neocortex, that hippocampus was more open as a place for Buzsaki to make his mark.
(By the way, I owe thanks to Uzi for pointing me to pay attention many years ago.)
TIME: both neocortex and neocortex are governed by powerful regular timing. Ideological people often say "The brain is asynchronous, no clocks at all, not like a computer, like a nice ODE." I was so happy when I started work at NSF, met the top neuroscience funder, and he said: "Neural networks? We have no interest in that kind of math, because they are all asynchronous. From REAL neuroscience (like Llinas) we know that the brain is fULL of high precision clocks. Until you show me models which show and account for clocks, I will not even look at it." That was fun -- but gaps between communities survive to this day. "
Werbos and Davis analyzed high speed deep recording data BOTH from hippocampus AND from neocortex (from Buzsaki's lab).
Of course we studied many of Buzsaki's papers first. He had a thorough review asking :"Where did that 4 cycles per second (theta) rhythm in hippocampus COME FROM? " He studied many possibilities exhaustively, including the possibility of clock control of hippocampus, but none of it worked. It ended as a mystery. But in MY model of decision-making neural networks (RLADP),
there is a 2-to-1 COUPLING between networks which measure value (which hippocampus is part of) and networks which learn prediction (neocortex), the known COUPLING between hippocampus and neocortex can explain the hippocampus rhythm, WHEN we have driving from the neocortex. For the neocortex, we have known for decades that there are pacemaker cells in the thalamus ("nonspecific thaalmus") which give the dumb pulse timing signals which go drectly to the central control points of all the giant pyramid cells of neocortex, the backbone of the neocortex. What's more, we measured the numbers. They do vary from animal to animal, and species to species, but in any animal it is HIGHLY PRECISE,
By the way, our paper TESTED the new, functional kind of model, with clocks, against the old (still popular) asynchronous types of model, and the new model wins hands down empirically.
They asked; "But WHERE is the pattern recognition.."?
In MANY places, at many levels.
Think of the neocortex as a kind of great forest. Actually, it is like the forest behind our house, where great trees rise from the dirt below ("layers V and VI") to the canopy above (layers 1 and II), BUT ALSO there are long strong wires up in the canopy level carrying power over long distances. The world of our awareness is basically what we "see" as the outputs of these giant pyramid cells, like where they are rooted in the dirt.. but under each tree is a strong root, the axon, which lagter rises up, branches and crosses long distances, up into the canopy. TOGETHER, these trees make up a PREDICTIVE MODEL of the visual inputs (and other assocaited sense inputs), and some of them have roots going all the way back to the thalamus. An integrative model.
BUT: here is where Freud comes into it. Some parts of Freud were much more scientific than most people know even now (despite Pribram's great efforts to explain). The giant pyramids, the big trees, give us the "ego" of his core theory of psychodynamics.
MODEL-BASED PREDICTION. But in the shadow of these trees, in various layers of this forest, are smaller cells which
can simply MEMORIZE or PHOTOGRAPH important past experiences. They offer a system of ASSOCIATION or MEMORY based forecasting, which fits well with Freud's "id." The "id" and "ego" work together to predict the environment. The ratio between id and ego can vary a lot, from person to person, time to time, and topic to topic. (And yes, I have addressed the math of optimal blending of the two, but more research is needed and important.)
They even asked: Is there a connection between the alpha/theta underlying rhythms and things like heartrate?
Of COURSE I mentioned Robert Kozma and our discussions, and Walter Freeman.
"It is less about heart rates and more about sniffing. The thalamus gives a very direct movie of the visual world, but also inputs highly processed versions of most other sense inputs. But there is also a powerful secret back door to get into the neocortex. Perfume makers even hire neuroscientists who study that backdoor, the SMELL system, which is more ancient than vision and can sneak into the brain bypassing the usual front door (thalamus). Actually, the olfactory system also is a major part of our sense of taste as well.
They asked: "More ancient than vision? How could that be?"
"Well, even for our most ancient mammal ancestors, like rodents, sniffing was bigger than vision. They didn't have our kinds of big eyes, But in fact, smell is really just our main CHEMICAL SENSE, and that goes all the way back to fish... "
"In truth, Robert and I have not done all the brain data work we could to DISENTANGLE the impact of sniffing rhythms and
nonspecific thalamus clock inputs. [More analysis of the Buzsaki data is one of the major new areas for research
we would wish for, and been blocked on.] But of course, the images on a movie screen are the result BOTH of the rhythm
of the projector AND the actual stories being projected on the screen."
Next they asked about the neuroscience of music. I cited discussions years ago with Judy Franklin
https://www.researchgate.net/
We agreed that human language STARTED with the mirror neurons we share even with monkeys, but got a real impetus from the "bushman dance" in those videos, which was ever so impressive: a system for TRANSFERRING certain important memories from a hunter to others in the group, including receptive states in the listeners, an important step in evolution. Formal grammatical languages came MUCH later than the word dances which emerged from those hunter-gatherers.
==========
So that's all. Thanks to any of you who got this far. I believe it is EXTREMELY important, to know how brains work in a functional sense, with higher learning capabilities, and new research is needed to fill in and extend, and prove, and get the word out.
But as I said, I do not have a real institutional affiliation even at this point (unless maybe as an associate of one of you).
=====================================
=======================================
In that discussion, I mentioned how Werbos and Davis (https://www.frontiersin.org/articles/10.3389/fnsys.2016.00097/full) showed how the CLOCKS work, moving from cerebrak cortex to hippocampus. How they are powered by pacemaker cells in the nonspecific thalamus, which output a dense tangle of fibers to the central control piints in the IDDLE of all those giant pyramid cells. (I often cite work by Purpura and Schebeil and Scheibel on what the anatomy looks like.)
They also asked about "where do the valuye, the utility function U and J and lambda come from?"
The LEARNED values, like J and lambda, and secondary reinforcement, come from the "limbic system," which I discussed at length in earlier papers. But that system is drive by the underlying values, the U, which mainly come form three places: (1) the coupled system of hypothalamus and pituitary gland, which monitor chemical flows in the blood (like low blood sugar or hormones) and give basic pain and pleasure input to the rest of the limbic system; (2) the coupled system of epitah;amus and pineal glad, just as important, but harder for neuroscientists to nail down because it includes response to social cues and things which it is harder to measure in wet labs; and (3) mammillary bodies, systems which use ancient algorithms to estimate value variables requiring some complex processing.
============
In a further LEF discussion, we discussed posting further information and a dialogue system at www.ffsii.org. We discussed language in more detail, important to people building dialogue packages.
I discussed how the ancient Bushman dances (similar to dances in other old tribes) allow a certain kind of rapt listener to ABSORB the experience of the dancer into the database of memories which the liatener can learn from. This direct and powerful system naturally leads to development of compact, lower cost condensed versions, like use of music and vocal utterances to portray a sequence of experience, either one specific memory or the dancer's internal condensed memory of a SET of similar experiences. The utterances become regularized, and lead to "word movies," the ancestors of asenetnces in the most ancient uninflected less formal languages. (I cited Sapir.)
We also discussed some modern languages where a more ancient, poetic "word movie" literature may exist side by side with more formalized, unnatural languages like modern English, where sentences are restricted to what follows thge kind of rules which Socrates and Max Weber would be proud of.
Many modern writers talk about the gap between the "conscious verbal" self and the vast, unaware nonverbal self. I commented that schisms and gaps vary a lot in human brains. The old split brain experinces are a great clue. Schisms between verbal and nonverbal areas, or between left and right brain, or between neocortex and emotions (like limbic system) and even between brain and soul are all
imperfections, of lack of full development, in human brains. The need for better education, K12 and beyond, to learn greater integration and empowerment of human minds is one of the most important grand challenges before us all today.
=================================================
=================================================
Further questions from a friend:
I have many questions, I start with four:1- What is the (bidirectional) relations, if any, between these brain clocks and our "illusion" of time?
The hard part of this question is to nail down WHAT IS "our illusion of time."
In truth, our ACTUAL experience, the CONTENTS, is a complex multilevel system (even just considering the mundane mammal brain). The universal learning mechanisms
are essentially unchanging, but WHAT we learn varies so much. MANY of the feelings we talk about tend to straddle multiple levels, and have many aspects. Pribram's little book on Freud, from Gill, does a great job of dissecting the MANY different real things discussed in Freud's writings on psychodynamics!
So -- which illusion of time?
The MAIN clock, the pacemaker cells of the nonspecific thalamus, just gives timing pulses.
No LABEL of what time it is.
When we say "time seems to be passing faster now," we usually mean that the CHANGES from moment to moment seem larger than usual. Than usual? SOME kind of time labelling is present in all the many kinds of memory we have. But LARGE CHANGES from alpha time pulse to alpha time pulse can be experienced as "time passing faster than usual," even without memory. (By the way, in a rat, that would include the rate of sniffing.)
Buzsaki's HIPPOCAMPUS is usually seen as the place WHICH PUTS LABELS on episodic memories (memories of sequences of "images" drawing on one or more senses). THOSE LABELS are responsible for how fast things seem to move when we recall an episodic memory.
Of course, neocortex itself is where we actually form actual images of actual clocks and watches, and our "experience of time" also includes feelings and images we experience as we look at actual clocks and watches.
The basal ganglia, which execute complex actio skills (like "open the door"), entail
various senses of time, ranging from frustration if it does not open fast enough, to
the sense of time about "before and after" the time it is opened.
By the way, everything I have said so far here addresses... the usual "conscious mind," the bulk of the brain. But lower parts of the nervous system are also important. Above all, the
CEREBELLUM (receiving orders from above, basal ganglia and other pathways output from neocortex) has its OWN cortex, a faster and simpler structure, but timing is certainly important in motor coordination as well. I sometimes compare the cerebellar system to a football player, and the upper brain as a coach. The football player is much faster and more coordinated, but has a kind of tunnel vision, not seeing the big picture or seeing far into the future. So there are lower clocks too. Llinas was the number one guy outside Japan in studying that system.
--------
**** IF we ever get a chance to get research to build on what we have learned, one
of the first simple tasks is to REPLICATE what Joshua and I did for one of Buzsaki's sessions
with the OTHER sessions. (That may include trying new spike sorting algorithms we have sketched out, algorithms to add annotations to the full database from Buzsaki.)
I expect that the same rat will show exactly the same UNDERLYING clock cycle as in our paper
(though Fourier analysis may yield different time units related to changes in sniffing and such).
BUT WHAT OF how the clocks behave, in neocortex and hippocampus, when the same rate is sleeping, deep sleep or dreaming? This would be an interesting thing to learn, and Buzsaki's data does include that kind of alternate session of brain recording. This is a case where a new empirical result would be totally new and important.
=============================
2- Can we change the pace of these clocks say by drugs/chemicals?
**** I do not know whether anyone has ever studied the outputs of the thalamic pacemaker cells (or other pacemaker cells) as changed by drugs or chemicals. For the general alpha and theta rhythms (like ouputs of cells in neocortex), I would EXPECT many such papers, but I haven't looked for that. Do humans or rats sniff more when on LSD?
By the way, when my wife Luda and I visited the German embassy years ago, they had an open ceremony to honor the German researchers whose work they liked the most in the Society for Neuroscience big meeting in DC that year. One was a woman who discovered some kind of smaller pacemaker cells, pulsing out higher frequency local rhythms used in local circuits within middle levels of neocortex, they said. I have ideas on how they fit, maybe analogous to cerebellar pacemakers. It is another area for future research. LOWER frequency pacemakers like alpha and theta are crucial to synchronizing highly recurrent networks where real time is especially important, over larger chunks of it.
But even local fast feedforward networks can benefit from precise synchronization as well, as is usually just assumed when we build such systems in engineering.
3- How do these brain structures work during the hallucinations induced by psychedelics?
I have been unusually strict in avoiding psychedelics for my whole life.
That resolve started in high school (when it was an issue in my class.) More precisely, from 1962-1964 I attended Lawrenceville school in New Jersey, where Turqi Faisal was a classmate. Not counting him...
I was told one year than in a class of more than 100, my close Chinese friend and I were the only ones who never even came near a joint of marijuana. Why was an outlier?
By then, I understood the neural networks of the human brain as an incredibly evolved system,
tuned TO WORK. I wanted mine TO WORK. Have you ever seen that sommercial where they first show an egg ("This is your brain"), and then crack it and fry it and say "This is what drugs do to your brain"? In my crazy old age, I actually believe that I broadcast that thought SO loudly that it helped create the commercial.
But tea was always there, and in the working world coffee and even times of highly controlled alcohol use..
In a visit to Udaipur in India a few years ago, we were showed a new article in Smithsonian magazine reporting the big change at NIH, allowing controlled experimentation on psilocybin. "Does it open up
the windows of the soul to seemore?" I cited a book by Annie Besant (Thought Forms, one of the books in Ghandhi's personal sacred shelves we visited later in Mumbai) STRONLY opposing that use of psychedelice. "Your windows will open NATURALLY if you develop the ability, in brain and soul, to make positive use of the information. Before that, you just hurt yourself. Just this morning, in a state of higher meditation in bed, I thought back to her and to that Daoist teacher I once met who said: "Enlightenment is easy. we could give you that in five minutes. The purpose of our intense long discipline is to allow you to cope and survive after that happens."
So maybe my main answer here is that I don't know, and don't really want to know.
At NSF, I always stressed "my goal is to understand how HEALTHY brains work. Disease and physical aberrations complicate things so much; I am happy to leave some of these issues to our friends at NIH, whom we work with at times but do not crave their turf. And we have enough work to do as it is, without diluting it."
Life being what it is, I HAVE learned some stuff about various chemicals, but it's not about clocks, so I won't elaborate here.
HOWEVER: the action of hormones and genes ISan important part of the normal healthy mammal brain, from mice to humans. These chemicals (and the old trio of tea, coffee alcohol) DO have implications for the interface of brain and soul as well. If you folks are interested, I should maybe someday forward what i sent our local friends responding to work from Harvard and Yale on brans and gender, and going further.
4- Finally, when you talk about prediction, is that in the sense of future times or as an estimation or reconstruction of an ongoing state such as with an observer or Kalman filter?
FOR NEOCORTEX, it is strictly pacemaker time that drives the error signal that implements the main function of neocortex. For the "outer loop", I am thinking of the SEDP equations (a simplified model, but correct in spirit I think) given in chapter 10 of the Handbook of Intelligent Control, a chapter I posted at www.werbos.com/Mind.htm. The prediction cells measured by Nicolelis predict the raw input cells
(actually, whisker tweak sensing cells, the "barrel system") in the thalamus, WITH a lag time he reports.
**** IF we could find Buzsaki type deep recording data for SUCH RATS, with tha;amus cells also tracked, we could show how precise the agreement is between the lags in the predictions Nicolelis reported, and the clock cycle which Joshua and I measured. In theory, maybe one of Nicolelis' rats
could be instrumented LATER ala Buzsaki.
*** The key point, in our understanding of how brains work, is that the timing signal which hits the middle of the giant pyramid cell, is similar in a way to an on-off switch. SEDP is a nonlinear generalization of the Kalman filter, so yes, it is a good analogy. The smaller cells in neocortex PERTURB it, as do changes in actual input images like sniffing, but the pacemaker is the basic rhythm.
Now that I think of, some kind of Kalman filter simulation could illustrate how periodicities (like sniffing) in the inputs can result in frequencies and emergent rhythms different from simple t/t+1 of the pacemaker.
I would like to encourage you to write these systemic/integrative brain findings in a more elaborate manner with drawings for illustration purposes; no math though unless necessary!
As I wrote this, I thought often of slides I have shown which do that. I included some in my 2009 paper in Neural Networks, for which I received the Hebb award.
No comments:
Post a Comment