My candidate structure is a filter cell in a synaptic matrix. See  "Sparse Coding of Faces ..." on my RG page.

There is no single cell storing an item. The information is stored in syntactical weights in the whole neo-cortex. 

Agree, with Wilfried. I would describe it as a distributed cell assembly (likely over several cortical areas and perhaps modalities) of connected pyramidal cells, in larger cortices likely grouped in functional minicolumns. 

I would say that the cellular structure of the neuron is an irreducible unit for memory and cognition. It is the base of a communication network, that in modern conceptual formulations is neither Boolean, Bayesian etc. It includes the self-reference and is hence denoted "Gödelian" including its Call Centre Poisson-like distribution.

The memory forming component, i.e. the spatio-temporal mnemonic entity, defines a structural constituent that incorporates spatial order in the chemical laboratory defined by the synaptic cleft. It is rooted in the genetic code as well as develops the higher order temporal program of the cell. The common denominator is their Poisson statistical distribution.

It goes without saying that this "Call Centre" property cannot work without e.g. environmental assistance, in particular the astroglial neurotransmitter-releasing Master Hub.

The irreducible formation of memory and cognition, makes it difficult to define a particular minimal responsible structure.

I partially agree with Wilfried. Artificial neurons

are often described with synaptic input weights

and one special value, the threshold. If the sum

of all inputs multiplied with their synaptic weights

exceeds the threshold, the neuron fires. Of course

the threshold isn't really necessary, it can be

expressed as an additional synapse coming from a

constant input, e.g. -1.

Real life neurons are more complicated.

A single neuron can detect temporal sequences.

http://www.eurekalert.org/pub_releases/2010-08/ucl-snc081110.php

http://www.eurekalert.org/pub_releases/2013-10/uonc-und102513.php

http://www.eurekalert.org/pub_releases/2013-10/ucl-sns102513.php

http://arxiv.org/abs/1511.00083

http://news.bbc.co.uk/2/hi/health/7151920.stm

Connectionists often avoid such discussions by

speaking of single units instead of neurons.

The code also plays a role. There are dense

codes like the binary code, or some other extreme,

the one-hot-code. Of course neurons can map

between these representations.

For example, there are the grandmother neurons.

One neuron encodes one symbol (or item).

http://www.newswise.com/articles/view/512626/

http://medicalxpress.com/news/2012-12-brain-cells.html

https://www.newscientist.com/article/dn17508-oprah-neuron-hints-at-nature-of-memory/

One the other hand, working memory seems to

have to do with brain waves. That means, thousands

of neurons contribute to a symbol (or item).

http://www.eurekalert.org/pub_releases/2012-11/niom-ibw110212.php

Regards,

Joachim

This might help with the answer? Researchers identify a specific mechanism by which the brain connects separate memories.

http://www.nature.com/nature/journal/vaop/ncurrent/full/nature17955.html

Memory function is not only based on the concrete stored item. There are also association fibers to emotional and other areas so we have an universe of different aspects of the stored item. All these aspects of a single stored item distinguish brain stored items from technical stored items which miss all these relations to other aspects.

There must be some kind of maintenance process. Otherwise, information stored in individual cells will slowly drift, leading to an overall memory distortion. Say, memory continues to be subconsciously replayed forever instead of being "recorded" as synaptic weights.

There is no way to know exactly (and not calculate) the smallest unit of memory in the human brain.

Every human brain is composed of approximately 1000 million neurons. Each neuron in turn creates approximately 1,000 connections with other neurons, resulting in more than a billion connections.

If each neuron save part of a memory or a full memory, this would be the smallest unit of memory, not the same organelles (mitochondria, nucleus, endoplasmic reticulum, ribosomes, etc.); ie it would complete neuron as a whole would function as the minimum memory unit.

According to clinical and experimental studies in humans (project human connectome) and in animal models, neurons operate in network and (synapses) are connected to other neurons, so that each contributes many memories while increasing exponentially storage capacity of the brain to something closer to 2.5 petabytes (1 Pb = 1 million gigabytes.

On the other hand, the minimum memory of computers equipped with artificial intelligence unit is the bit (binary unit: 0 and 1). Which they are far from being compared to the human brain. Probably a neural network of 100 or 1000 micro-computers powerful running in parallel, are the equivalent of one human neuron.

regards

Sincerely

Dr. Jose Luis Garcia Vigil

From psychology we know that the transfer rate

at which memories are stored in long term memory

is at most about 1 or 2 bits per second.

That means in a lifetime of 70 years we can

only store about 250 to 500 megabytes, which

fits on a CD.

A thought experiment. Imagine a sheet of paper

with random characters, say 80*25. How long

does it take to learn the page by rote?

Hours, if the above 1 or 2 bits per second

are correct. Which is in accordance with what

can be observed.

Regards,

Joachim

Joachim,

I agree but random characters are nonsense. It shows only the transfer rate of pure information and not the quantity of information learned and stored in a whole life.

If we learn in everyday situations we learn not pure information but relations between stored items or complex data. Therefore you will have some problems with the capacity of a single CD.

Wilfried,

Random characters are the only way.

The opposite leads to curious anomalies.

How fast could you learn the Kyrillic

alphabet if you already know 90% of it?

Very fast. The correct way of measuring

the learning speed is to start off from

scratch. No reference to something

already learned should be possible.

The worst case of input should be

applied.

There is the Hutter Prize. A certain part

of the Wikipedia must be generated from

a compressed file and a program.

https://en.wikipedia.org/wiki/Hutter_Prize

https://de.wikipedia.org/wiki/Hutter-Preis

Now what would you think, what's the

smallest possible size of the compressed file

that a decompressor can expand to the

Wikipedia? Of course it's zero. Then the

program code would have to be larger.

Therefore in the Hutter Prize, the size

of the data and the size of the program,

are added, only their sum counts.

Not the data alone.

In psychological experiments where the

learning rate is to be determined,

the possibility of compression should

be avoided.

Regards,

Joachim

Joachim,

do algorithm and neuronal nets co-relate in data processing?

It is the question of anticipation - one of my favorite questions!

Here is a try to compare the human mind

and computers. The basic data type of the

mind is a symbol. It can mean a word, a

smell, a sound, a taste, a touch, any

perception. A symbol can reference something

else. For example, a sound can stand for

a smell, and a smell can stand for a word.

A symbol is simply a reference.

Computers can also work with something

simular: with pointers that reference

something in address space.

Both computers and the mind have a

similar bottleneck. All conscious

processing takes place in working memory,

all arithmetic processing takes place

in the CPU.

Regards,

Joachim

Memory can not be reduced to a minimum level. The memory is based on the integration processes in 3D format at the molecular, sub-cellular, synaptic, cellular, neural network levels with the involvement of glial cells and matrix.

If by a memory you mean a snapshot of some event in the past that flashes into your consciousness, I think it consists minimally of activity in a group of neurons that together represent all of the aspects of that snapshot -- something like a "cognit", to use your terminology. Once activated, the connections between those neurons and others can evoke snapshots of what preceded and followed the remembered snapshot in the remembered event.

The minimal neural structure that stores an item of knowledge must be activated preconsciously and bound with the output of other preconscious memory structures before it is consciously experienced.

Joachim,

your answer is quite good, but does not answer my question.

Dear Wilfried! You have demonstrated an example of the top in tact response. Vladimir

Many thanks to all for very interesting answers.  Although you treat it respectfully, I admit that the question was a bit misleading, if not unreasonable, because the term “minimal” implies that the structure in question is measurable by present means, which by its very nature may not be.  But the question was not aimed at identifying a quantifiable entity but at establishing a principle that seems to hold for old cortex as well as new cortex.  In neuroscience, as in any other science, the theoretical principle is key to progress (“without a theory the facts are silent,” Hayek said).

            In my opinion, the correct answer to my question is: “none of the above.”  Several of you, however, came close to the most correct answer, which in my view is what I call a cognit: a network of neuronal assemblies representing the sensory and/or motor qualities of a discrete life experience, synaptically linked together by their simultaneous occurrence in that experience.  

           Cognits are of many sizes, depending on the complexity of the memory or item of knowledge they represent.  The larger ones are widely distributed, straddling several areas of the cortex.  Cognits overlap one another profusely and share common nodes or features, whereby one neuronal assembly can be part of many memories or items of knowledge.  They are all inherently plastic, susceptible to growth and recombination with learning and life experience.  Their most important structural properties are: (a) hierarchical organization, (b), nesting, and (c) sharing of connective links and neuron assemblies.

           The evidence for the structure and dynamics of cognitive networks or cognits comes from the study of the activation of cortical areas and cells in well-structured behavioral conditions (review in Fuster, 2009):

1.  Working memory (WM) tests.  WM is the temporary activation of a cognit for attaining a behavioral or linguistic purpose.  The electrical and imaging signals of that activation confirm the distributed character of cognits and their associative (synaptic) character.  Vivid evidence of the latter is the activation, during WM, of neuronal assemblies in dispersed cortical regions that retain cross-modal information, that is, information encoded by learning the use of stimuli of more than one sensory modality (e.g., vision and hearing, vision and touch).

    Gibson,J.R. and Maunsell, J.H.R. (1997). Sensory modality specificity of neural activity related to memory in visual cortex. Journal of Neurophysiology 78: 1263-1275.

    Hadjikhani,N., and Ronald P.E. (1998). Cross-Modal transfer of information between the tactile and the visual representations in the human brain: A positron emission tomographic. Journal of Neuroscience 18: 1072-1084.

    Fuster,J.M. et al. (2000). Cross-modal and cross-temporal association in neurons of frontal cortex. Nature 405: 347-351.

     Zhou, Y.D. and Fuster, J.M. (2000).  Visuo-tactile cross-modal associations in cortical somatosensory cells. PNAS, 97:9777-9782.

2.  Formation and association of cognits by temporal approximation of their contexts.  In rodents, where spatial navigation has high survival value and where the hippocampus integrates cognitive functions that in the primate will be integrated by the neocortex, two important cognit properties have been upheld: (A) Time is of the essence for the formation of synaptic links between cognits, and thus for the formation and generalization of higher-level cognits. (B) Within those hierarchically higher cognits there is substantial overlap of the constituent cognits.

    Fuster J.M. - Cortex and memory: emergence of a new paradigm (2009). Journal of Cognitive Neuroscience, 21:2047-2072.

    Cai, D.J.  et al. (2016).  A shared neural ensemble links distinct contextual memories encoded close in time.  Nature, May 2016.

3.  Of course (re. Arnold’s answer), all those changes can occur unconsciously or preconsciously.

Cheers and kind regards.

Joaquin   

Dear Joaquin! Thank you. "Without a theory the facts are silent". Vladimir

Wilfried,

a Turing machine can simulate any other computing machine,

such as the brain. And the brain can simulate a

Turing machine as well.

Neurons work analogously. Real numbers can carry an

infinite amount of information. That is not a case for

making a difference between computers and the brain

since it is only necessary to simulate neurons up

to a certain precision. This, because neurons

themselves are only working up to a certain

precision, and a single neuron can always die,

a single neuron does not much influence the result

of a thought.

A brain can simulate a Turing machine. But extremely

slowly, and noone does ever think this way. The total

state of a Turing machine is made of the TM's internal

state of the final automaton, and the state of the

tape between leftmost and rightmost written character.

The brain is able to remember long sequences of

digits, think of the number pi. But this only goes

over the long term memory. Doing a computation in mind

this way would produce a, in the worst case

exponentially with time increasing, number of long

lists of digits. It would work, but take eons

and produce a lot of garbage lists.

But there are other machine models. Some that

can be faster simulated by the mind. One is

forward chaining production systems. A production

system consists of facts and rules. Facts are

tuples of symbols like (father mike tom) and

(father tom jim). Rules could then be:

Forall ?X, ?Y, ?Z: if (father ?X ?Y) and

(father ?Y ?Z) then (grandfather ?X ?Z). Note that

a small amount of such tuples fits into working

memory. The above rule would then infer:

(grandfather mike jim).

In my opinion forward chaining production systems

are very close to the human symbolic thinking process.

Regards,

Joachim

These are the microtubules. According to a theory

named Orch OR, they can perform quantum computations.

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3470100/

Related articles:

http://arxiv.org/abs/quant-ph/0204021

http://phys.org/news/2015-08-neural-qubits-quantum-cognition-based.html

http://arxiv.org/abs/1508.05929

http://arxiv.org/abs/0802.1835

http://phys.org/news/2011-02-wustl-physicist-debates-quantum-mind.html

http://www.newstatesman.com/society/2011/05/quantum-theory-consciousness

One of the problems with quantum computation

is that states can't easily be refreshed.

When a qubit is read out, it makes a choice

and is no more.

Regards,

Joachim

Networks are not mechanisms. In the brain, neuronal networks enable neuronal mechanisms to intercommunicate in useful ways. The minimal neuronal mechanism for storing a long-term memory must have the capability to recognize and label distinct patterns of neuronal activity.

Arnold,

True.  The cognit has that capability to be recognized by association.  Thus the network is activated and put to use.  Cheers, Joaquín

Joaquin: "The cognit has that capability to be recognized by association."

"Association" loosely covers all kinds of events, but we want to know how it works. Is the cognit something like the synaptic matrix in "Sparse Coding of Faces ..." on my RG page?

Professor Trehub,

I'm not familiar with the synaptic matrix you refer to, but here's what I think Professor Fuster means by association between cognits. As I understand it, a cognit is a group of neurons active during a mental experience. For example, right now I'm looking out my front window and I see my parked car, a pole for a basketball hoop, some houses and trees in the distance, and a pickup truck driving toward the cul-de-sac across the way. All of these things must be represented by activity in what is probably a huge number of neurons in my brain, and those neurons constitute the cognit. When I am no longer looking out the window, but am thinking about what I saw, there is presumably activity in at least some of those same neurons. Activity in the neurons active while I'm thinking about the experience as well as in those active during the experience itself can activate other neurons that might not have been part of the original cognit if the synaptic connections between them are strong or numerous enough. Those other neurons could represent memories of related experiences, and activity in them could activate yet more. Thus the neurons active in the original cognit can, through synaptic connections to other neurons, cause cognits sharing at least some of the aspects of the original experience to become active.

John,

I agree with your account of the cognit. But in order for the cognit to work it must have neuronal mechanisms that can selectively learn, recognize, store and recall relevant objects and events. That is why I asked if the synaptic matrix might be an essential part of any cognit.

Dear John,

I cannot say it any better.  But Professor Trehub is right too.  The cognit is a synaptic matrix.  And hierarchically higher cognits, that is, those that abstract and generalize (semantic memory among them), are subject to the connectivity principles of sparse coding and “small world.”  Cognits are formed, stored and retrieved by association, i.e., by synaptic activation, facilitation, and modulation.  These are the essence of the mechanisms that he is referring to, and therefore the essence of association, which selects neurons and cognits.  Thus here the term “association” is both structural and dynamic.

Cheers, Joaquín

Joaquín said: "I cannot say it any better. But Professor Trehub is right too. The cognit is a synaptic matrix. And hierarchically higher cognits, that is, those that abstract and generalize (semantic memory among them), are subject to the connectivity principles of sparse coding and “small world.” Cognits are formed, stored and retrieved by association, i.e., by synaptic activation, facilitation, and modulation. These are the essence of the mechanisms that he is referring to, and therefore the essence of association, which selects neurons and cognits. Thus here the term “association” is both structural and dynamic."

Thanks for your compliment, Joaquín, and thanks for posting the question that initiated this interesting discussion. I agree that the synapses connecting the neurons in a cognit are part of the cognit. They enable the entire cognit to become active, evoking the mental experience it represents, when only a part of it was initially activated. I suppose those synapses become strengthened through some sort of Hebbian-like mechanism when all of the parts of the cognit are active during the organism's first encounter with the mental experience the cognit embodies.

What is particularly interesting to me, however, is how the neurons of a given cognit become synaptically connected to the neurons of the cognit representing the appropriate response to the activity in the first one. In other words, how the organism learns what the first cognit means, how it learns what to do or think after having the mental experience the first cognit represents. My view is that the strengthening, weakening, and rewiring involved in establishing those connections are driven by a mismatch between a cognit (or cognits) representing what the organism "wants" (some goal it has) and a cognit (or cognits) representing what the organism "has" (its perception of its current situation). I would appreciate any comments you and others might have about this.

Also, do the "hierarchically higher cognits" you mention above consist of a kind of concatenation of hierarchically lower cognits, or are they completely separate sets of neurons?

Thanks

John

John,

 For a theoretical model relevant to your questions, see "Learning, Imagery, Tokens and Types: The Synaptic Matrix" and " Self-Directed Learning in a Complex Environment" on my RG page.

Arnold Trehub: "For a theoretical model relevant to your questions, see "Learning, Imagery, Tokens and Types: The Synaptic Matrix" and " Self-Directed Learning in a Complex Environment" on my RG page."

Thanks.  I will take a look at those.

John

Hi John

          “I suppose those synapses become strengthened through some sort of Hebbian-like mechanism when all of the parts of the cognit are active during the organism's first encounter with the mental experience the cognit embodies.” Indeed, a Hebbian mechanism of synaptic modulation by temporal coincidence is indispensable in that first encounter as well as for all subsequent phenomenal experiences derived from it.

         “`In other words, how the organism learns what the first cognit means, how it learns what to do or think after having the mental experience the first cognit represents.” This is incomprehensible if you take cortex and its cognits out of the Perception–Action Cycle, the cybernetic mechanism that regulates the adaptation of the organism to its environment.

J.M. Fuster and S.L. Bressler – Cognit activation: a mechanism enabling temporal integration in working memory.  Trends in Cognitive Sciences, 16:207-218, 2012.

        “My view is that the strengthening, weakening, and rewiring involved in establishing those connections are driven by a mismatch between a cognit (or cognits) representing what the organism "wants" (some goal it has) and a cognit (or cognits) representing what the organism "has" (its perception of its current situation). I would appreciate any comments you and others might have about this.” Here again you need the PA Cycle and the matching and error-correction mechanisms of the medial--cingulate--prefrontal cortex within the Cycle.

        “Also, do the "hierarchically higher cognits" you mention above consist of a kind of concatenation of hierarchically lower cognits, or are they completely separate sets of neurons?”  No, the cognits of higher order nest those of lower order.  So in part they share some of the same neurons.  But there is some heterarchical connectivity between cognits of different order to make up those of the higher order.

Best,

Joaquín

 Hi Joaquín

Thanks for those comments, and the reference.  Have you ever heard of perceptual control theory?  Here's the seminal reference: Powers, W. T. (1973). Behavior: the control of perception. Chicago, Aldine Pub. Co.

John

John,

How does perceptual control theory deal with human imagination? You might be interested in reading "Set Point and Motive: The Formation and Resolution of Goals" on my RG page.

Hi John,

I have just ordered Powers' book.  From what I read about it, I gather his theory comes very close to my "Perception-Action Cycle," or vice versa. Of course, both have their origin in systems theory, Uexküll, and Bertalanffy.  It sounds fascinating.  Thanks for bringing it to my attention.  Joaquín 

Arnold: How does perceptual control theory deal with human imagination?

Perceptual control theory (PCT) imagines the brain as a hierarchy of control systems, like thermostats, with control systems at higher levels of the hierarchy sending "reference signals" (like the temperature setpoint on a thermostat) to "comparators" that compare the reference signal to the "perceptual signals" (like the temperature of the room) coming from lower levels.  If there is a mismatch between the reference signal and the perceptual signal, the control system generates an "error signal" that causes some action--picking up a cup of coffee, for example--that makes the perceptual and reference signals match.  If for whatever reason the perceptual signal can't be made to match the reference--maybe the cup is on the other side of a transparent wall, for example--then "reorganization" of the hierarchy occurs until the error signal is reduced--perhaps by figuring out how to open the door in the wall and reach through for the cup.

In the original formulation of PCT by Bill Powers (Powers 1973, see my post to Joaquín), there were nine levels in the hierarchy, ranging from intensities of sensory signals to "system concepts" like what kind of person you think you are.  Perceptual signals get sent up the hierarchy and reference signals get sent down, sometimes being stored along the way. 

To answer your question, PCT says that imagination occurs when perceptual signals are prevented from percolating up the hierarchy at some level, and the hierarchy is effectively experiencing memories in the same way as it would experience perceptions.  I hope that is not too obscure an explanation.

Joaquín: I have just ordered Powers' book.  From what I read about it, I gather his theory comes very close to my "Perception-Action Cycle," or vice versa.

Yes, there are some similarities between your P-A cycle and PCT.  I'm looking forward to continuing this discussion.

Thank you very much, John

John,

it seems to me that the concepts of reference signal in PCT and set point in my cognitive mechanisms are very similar. Would you agree?

Arnold: it seems to me that the concepts of reference signal in PCT and set point in my cognitive mechanisms are very similar. Would you agree?

I need to read your set point and motive chapter much more carefully, but after a quick look, the concepts of reference signal in PCT and your set point certainly do seem similar.  (I even used the term “setpoint” in describing the reference signal in a previous post here.)  And your description of eliminating “disequilibrium” to bring perceived events in the real world into agreement with a “secular goal” sounds a lot like what I was saying in a previous post about making what an organism “wants” match what it “has”.  The important question is, how does that actually happen? 

By the way, you might be interested to know, if you didn’t already, that a 1971 Science paper of yours, “The Brain as a Parallel Coherent Detector”, is referenced in Powers’s book on PCT.