Brain Building - Q1. Define Intelligence

[neubie]

You simply cannot make a decision without knowing if something is good or bad

I am not entirely sure about this. If we talk about higher degree of intelligence then I can agree to that. But if we are talking about basic intelligence such as when I look at a stop sign from various angles to determine if it is a stop sign, I don’t believe I need emotion. Don’t you agree?

Having said that

some element of judgement is part of the definition of intelligence

I believe you are right that one of the elements should be “Making Decision” which I think I have an idea how the brain makes decision but I won’t discuss the “how” in this posting and will leave it til later. I want to focus on coming up with the fundamental list of elements to define intelligence first.

[neubie]

I like to think of isocortex as a database of sorts, where indices of data entries are sparse partial representations that invoke attractor dynamics throughout cortex, which are synced together because all things in the database were learned through joint sensory agency and interaction. Once stored, you can think of the isocortex as a simulation, where not only can you match external sensory perception to internal representations, but you can invoke internal representations at will. All representations are massively associative, so you can easily travel from concept to concept via “train of thought” associations.

I have my own understanding on this and my understanding matches what you just said. Except I am not really sure how we “can invoke internal representations at will” and what is the trigger mechanism. To me, that is a very miraculous mechanism how we “can invoke apple”.

We (Numenta) think this model in isocortex is a fundamental component of intelligence

So can we at least conclude what this model is:

1. Perceiving Information => “joint sensory agency” and “invoke internal representations at will”
2. Memorizing Information => “a database of sorts, where indices of data entries are sparse partial representations”
3. De-memorizing Information => “a database of sorts, where indices of data entries are sparse partial representations”
4. Predicting Information => “All representations are massively associative, so you can easily travel from concept to concept via “train of thought” associations”

Still not sure about location in space…

[neubie]

If such (unfortunate) person has any senses (hearing, touch), then there is hardly any doubt a location framework will be produced to structure all those inputs.

So for this unfortunate person since he/she has no limbs this person cannot touch. If just hearing, how does a location framework work? If it is about the left and right ears, what if one of the ears has hearing loss? Does this person still have intelligence? Would there still be hardly any doubt a location framework will be produced to structure all those inputs? And if a location framework is produced, is it just a model formed by the intelligence and not a fundamental element of intelligence?

[neubie]

If I am a one celled organism I need to know where is the best food at this time of day and where is safe at this time of day

I don’t necessary disagree location is important, but I am wondering if it is essential. Similar argument to emotion, I think it is important for high order intelligence, but I don’t think it is an essential element.

If place is important, does that mean the single cell organism has intelligence? And I am using my example on this again for the super unfortunate person born blind without limbs and unable to move his/her neck, would this person have location model? and is location model important to this person? And if not can we conclude this person does not have intelligence. Please let me know your thought on this. Thanks.

[neubie]

I believe that “observing an apple” and “think apple” being distinguishable in this sentence highlights that they are perceived (and thus represented) in not exactly the same way

Agree on this.

They share some overlap in the neurons involved

Not sure if neurons actually represent the information or if the synapses do, but that would be something I would like to find out on a separate post.

Now do you think we can at least conclude that:

1. Perceiving Information => “observing an apple” and “think apple”

is one of the fundamental list of elements in defining Intelligence?

[Bitking]

So can we at least conclude what this model is:

1. Perceiving Information => “joint sensory agency” and “invoke internal representations at will”
2. Memorizing Information => “a database of sorts, where indices of data entries are sparse partial representations”
3. De-memorizing Information => “a database of sorts, where indices of data entries are sparse partial representations”
4. Predicting Information => “All representations are massively associative, so you can easily travel from concept to concept via “train of thought” associations”

All in the cortex? No - I will not agree that we can park any of this exclusively in the isocortex.

1. What perceives? Just passing through the sensory apparatus really can’t be called perception. Recalling existing memorized fragments and standing that up to build a representation in maps of the cortex is actually close to perception but it misses the involvement of subcortical structures in the perception of emotional import.
   Mixing this with recall in this item is not appropriate as this should be grouped with using the indexes to recall representation below.
2. having “surprise” trigger local learning is the closest to a pure isocortex function but it is very likely that this also involves the thalamus.
3. I am not sure what you intend with”dememorizing” but having a partial memory (index in your terms) auto-complete to the nearest match reactivates the memory in the exact place where it was first experienced - Déjà vu. There is reason to suspect that this memory is shared with connected maps as part of normalization of memories during the sleeping process but I don’t think that this is what you are looking for. The feedback path through connected maps surely involves the thalamus but the driver for this has to be subcortical structures.
4. This bit, thinking, involves recall and re-perception in the isocortex but it also involves the thalamus, hypothalamus, HC/EC complex, and very likely the cerebellum. The cortex, by itself, is purely a passive structure that receives some input and either matches it or learns it if it is novel, with the attendant output of bursting to signal the novelty. Perception drives the cortex at one end and subcortical structures (primarily the hypothalamus and cerebellum) drives it from the other end.

By the time you get to these very high level concepts you are very far away from the basic cortical column computation and have moved on to the emergent behavior of large ensembles of hardware. As such, the are not going to be the pure functions you are looking at for your foundation.

As unpleasant as you may find this - Numenta has been working at the same goal as you but seem to have discovered that the cortical column computation is the fundamental unit and this is combined in useful ways that result in these emergent properties that you are trying to describe.

[OhMyHTM]

When writing a program, someone will first perform a system analysis, abstracting the interface, and some people completely do not know what the interface is.
It doesn’t matter, let’s try to figure out all the details of the brain, then we may find that we absolutely don’t need “emotions”, we don’t need “apple”… All we need is a “laundry list”.

[Bitking]

You go first. What is the list?

[OhMyHTM]

@Bitking If you want to know what a list is, forget about the cortex, forget about the papers, and think about it from the perspective of an information system.

[Bitking]

No can do. The only example of actual intelligence I can point to is based in biology.
The deeper I dig into this the more I am driven back to that basic fact.

I have never seen an approach that was based on “folk psychology” that extract the “laundry list of features” that ever really went anywhere. Not for lack of trying - there have been many attempts.

Murray’s elusive g factor is an emergent property of parts hooked up the way the brain is constructed.

Of course, I am ready to be proven wrong so do your best with a working example.

[OhMyHTM]

okay, you are not wrong.
I have to borrow a picture from Jeff:

[bkaz]

I agree that definition is crucial, but you will never have everyone on board :). Yes, it’s all about prediction, which includes both what and where: location. But working from the first principles is mutually exclusive with imitating the brain, which is what folks here try to do. I do the former, not the latter, here is my take: http://www.cognitivealgorithm.info

[Paul_Lamb]

Not sure if neurons actually represent the information or if the synapses do, but that would be something I would like to find out on a separate post.

Since the synapses receive their input from presynaptic neurons, this seems a bit like a “chicken or the egg” question to me… both components are a necessary part of the system. It just tends (for me at least) to be conceptually easier to refer to the active neurons versus the synapses that they are transmitting into when talking about “representations”.

[neubie]

Definitely worth discussing more on this. The reason why i think the representation is in the synapses is because (could be total bogus):

1. no of neurons is only roughly 100B, but each neuron roughly has 7000 connections to other neurons. By judging how much we need to remember, I think we need more than 100B.
2. synapses has built in strength control through neurontransmitters, that allows how strong of the representation as a conclusion through association
3. synapses can grow and be pruned, that allows memorizing and de-memorizing.

But I think there are researches showing not neurons/synapses store representation. This is definitely a bigger topic and will create another post later on after I finish with this one.

[neubie]

but you will never have everyone on board

Wooo everytime I hear the word “never” it makes me excited. That means if somehow we can get through this the list will be a very strong one. Will take a look at your paper and learn more. Thanks!

[Falco]

(…) I am not really sure how we “can invoke internal representations at will” and what is the trigger mechanism. To me, that is a very miraculous mechanism how we “can invoke apple”.

@rhyolight believes in free will. (I know I know). We’ve had a very… very long thread about this somewhere here. (Don’t mind the thread title).

[neubie]

No can do.

Love it when I hear “no can do”, that makes me excited.

I am a newbie so I don’t really know what folk psychology is. But I get that you thought I didn’t come from a biological perspective to look at this issue. Although I am a software technologist, but biology was my stream in my A Level and virology was my favourite subject. I do love biology. For the same reason, that’s why I like “On Intelligence”.

Now maybe I will switch my question to another angle to align our perspectives. In your opinion, what is the list of the most basic functions the brain (not just our brain) possesses that allow us to conclude the brain has Intelligence?

I believe these are the one you listed on another post:

What is the Basic Cortical Column Computation?

• Pattern matching (by pattern recall)
• Novelty detection (by pattern recall mismatch)
• Pattern learning (by novelty triggering learning)
• Anomaly signaling (by sequence mismatch mechanism)
• Sequence learning (by anomaly signal triggering learning)
• Prediction (by sequence learning)
• “Memory” (by the collective action of these mechanisms?)

I assume you would agree if something can perform the above, then that something can be said to have intelligence? Correct?

[Falco]

what if one of the ears has hearing loss? Does this person still have intelligence? Would there still be hardly any doubt a location framework will be produced to structure all those inputs?

So with only hearing coming from a mono location in principle that person should still be able to learn language, right? Well, it turns out that even in the isocortical regions for language there are at least two layers that produce grid structures (6b and 5b, and possibly 6a). It is not clear yet (at least to me) how this works, but it is more than likely that all this hardware is used to structure the vast body of information required for understanding language.

This is as fundamental to intelligence as it gets.

Not sure if neurons actually represent the information or if the synapses do, but that would be something I would like to find out on a separate post.

I think you need to look a little deeper into HTM theory. You have some catching up to do.

[Bitking]

I assume you would agree if something can perform the above, then that something can be said to have intelligence? Correct?

No. That is a basic logic fallacy. An example should help illustrate the point - a computer is built using logic gates, but not every device built using logic gates is a computer.

Coming back you your idea that you can tick the boxes of a laundry list and POOF you have intelligence … the part that is going to be very hard to define - the secret sauce - is how the part are arranged. You can have all of the right element present but they have to be Just So or you did not win. Defining this arrangement is the bit that trips up the laundry list approach.

The cortical column computation seems to be sufficient to build intelligence in one part of the brain but that is just the starting point. There is still a lot of theoretical heavy lifting left undone.

I agree that listing the parts is a great starting point (which is why I put up the cortical column computation post myself) but that really is the starting point on a very long road and I do welcome fellow travelers. I am not trying to discourage you so much as trying to help you see what the shape of landscape we are traveling in.

[Paul_Lamb]

no of neurons is only roughly 100B, but each neuron roughly has 7000 connections to other neurons. By judging how much we need to remember, I think we need more than 100B.

Yes, just pointing out that you need both to do anything useful. Populations of neurons form semantic representations when active around the same time, and the synapses are what connect those representations to each other in a virtually infinite number of useful ways.

• synapses has built in strength control through neurontransmitters, that allows how strong of the representation as a conclusion through association
• synapses can grow and be pruned, that allows memorizing and de-memorizing.

Definitely. The synapses being the part that actively changes (while numbers of neurons much less so), they are where memories are being laid down. Just pointing out the fact that groups of synapses by themselves can’t do much without populations of neurons on either end to encode semantic meaning. They are both necessary parts of one system.