Is it possibly to functionally transfer knowledge from a neural network to another?

[Buckaroo Banzai]

I was wondering if you can even isolate bits of knowledge and copy just those, functionally, to another working network and integrate it with a different body of knowledge, or whether any given specific knowledge is too ingrained and highly contextualized into their original networks, and possibly even nonsensical to another, if just copied.

In other words, if you could "upload" arbitrary knowledge like it was done in the Matrix movies, but leaving aside the whole problem of how to artificially induce all the synapses and whatnot.

I don't know fundamentally anything about this subject.

 

[Michael 345]

I was wondering if you can even isolate bits of knowledge and copy just those, functionally, to another working network and integrate it with a different body of knowledge, or whether any given specific knowledge is too ingrained and highly contextualized into their original networks, and possibly even nonsensical to another, if just copied.

In other words, if you could "upload" arbitrary knowledge like it was done in the Matrix movies, but leaving aside the whole problem of how to artificially induce all the synapses and whatnot.

I don't know fundamentally anything about this subject.

If you mean extracting "thoughts" from one brain and reverse the process to insert them into another brain NO

The current method of teaching works with vastly vastly different levels of success depends on the myriad of variables. There is variation with methods and equipment which tend I think to help LEARNING and a few to HELP teaching

The art is to slip each into into the slots where they would produced the maximum benefit

 

[Write4U]

I would think the main problem is that no two brains are wired the same and/or operate the same way.
Perhaps something similar to trying to directly transfer information between two computers with different wiring and operating systems.

However at a very fundamental level, it might be possible, if the two brains were empathic. This phenomenon can already be demonstrated as part of the mirror neural function, where we whince when we see another person getting hurt. This is a really interesting phenomenon. Why should the person who is not getting hurt, produce the same brain activity and as the person who is getting hurt.

Apparently this ability is a result of millions of years of evolutionary experience and is already present in at least several species of monkeys.

So, at a very fundamental level it might be possible, because we can already do this by observation alone, so it seems we share this neural system in a general way.
But that is different from transferring knowledge which is not related to specific emotional responses.
As the Gredeback lecture shows, this ability of the mirror neurons is very "context specific".

 

[someguy1]

The subject line doesn't match the question. A neural network is a conventional program running on conventional hardware. You can certainly take a memory dump and load it onto another computer and have it pick up where the first one left off. It's trivial.

The assumption that a mind is a computation is wrong; but even if you accept it, it's a completely different question than the one about neural networks. Neural nets are just programs. All the latest AIs are just conventional computer programs, no different in principle than the Microsoft Windows solitaire game.

 

[Write4U]

The subject line doesn't match the question. A neural network is a conventional program running on conventional hardware. You can certainly take a memory dump and load it onto another computer and have it pick up where the first one left off. It's trivial.

The assumption that a mind is a computation is wrong; but even if you accept it, it's a completely different question than the one about neural networks. Neural nets are just programs. All the latest AIs are just conventional computer programs, no different in principle than the Microsoft Windows solitaire game.

Well, you will need to read the OP question again and while you are at it check out what a mirror neural system (network) is. The question, as I understand it it is, if one could dowload human knowledge, the cognitive part of the brain (the mirror neural network) and translate it into another human mirror neural system, i.e append it to the existing knowledge of someone else's brain, such as was done in the movie the Matrix.
https://en.wikipedia.org/wiki/Neuron
and
https://en.wikipedia.org/wiki/Artificial_neural_network

 

[Write4U]

To continue the discussion,

ANN, or Artificial_Neural_Network,

Artificial neural networks (ANNs) or connectionist systems are computing systems inspired by the biological neural networks that constitute animal brains. Such systems learn (progressively improve performance on) tasks by considering examples, generally without task-specific programming. For example, in image recognition, they might learn to identify images that contain cats by analyzing example images that have been manually labeled as "cat" or "no cat" and using the results to identify cats in other images. They do this without any a priori knowledge about cats, e.g., that they have fur, tails, whiskers and cat-like faces. Instead, they evolve their own set of relevant characteristics from the learning material that they process.

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

 

[billvon]

I was wondering if you can even isolate bits of knowledge and copy just those, functionally, to another working network and integrate it with a different body of knowledge, or whether any given specific knowledge is too ingrained and highly contextualized into their original networks, and possibly even nonsensical to another, if just copied.

The problem here is that neural networks do not contain "bits of knowledge." They have connection strengths, which taken in total result in certain behaviors.

Take a common example - a neural network used to recognize shapes from a camera. There is no one section that recognizes "apple" vs. "orange." In multilayer neural networks you can sometimes spot trends (i.e. the first level shows a lot of edge-detection activity) but you can't just trace out the connection strengths and get the recognition "function" for an apple.

 

[someguy1]

ANNs have a lot of popular mysticism about them. In fact they are entirely conventional computer programs, implemented on conventional hardware. This point appears to be insufficiently appreciated. There are no magic neural net computers. ANNs are a very clever way of organizing perfectly conventional computations. Many ANNs aren't even Turing complete, though some are. Any ANN can be executed (slowly of course) by a person using pencil and paper. That's why I said you can just copy over the ANN's memory state from one computer to another.

Of course how the mind is implemented by the brain is a deep mystery. That's why there are two entirely separate questions here: one about ANNs, and the other about minds and brains.

 

[iceaura]

Any ANN can be executed (slowly of course) by a person using pencil and paper. That's why I said you can just copy over the ANN's memory state from one computer to another.

But you won't know what you're doing, as far as transfer of the thinking ability.

The mystery of the "implementation" of mind will remain, only now it's a mystery of what a computer is doing.

You can in principle duplicate the hardware and snapshot software status of any brain, but that doesn't tell you what you need to know.

When the Go computer that has reached 9-dan professional rank plays a game, its moves are analyzed as consequences of its thinking - back-engineered, in a sense - just as a human player's are. One could of course download a copy of its state, in principle, at the moment it made its decision, but that wouldn't tell you what you want to know. At best it would provide informative hints for analysis, like a snapshot from a brain scan of a thinking person.

 

[Write4U]

But you won't know what you're doing, as far as transfer of the thinking ability.

The mystery of the "implementation" of mind will remain, only now it's a mystery of what a computer is doing.

You can in principle duplicate the hardware and snapshot software status of any brain, but that doesn't tell you what you need to know.

When the Go computer that has reached 9-dan professional rank plays a game, its moves are analyzed as consequences of its thinking - back-engineered, in a sense - just as a human player's are. One could of course download a copy of its state, in principle, at the moment it made its decision, but that wouldn't tell you what you want to know. At best it would provide informative hints for analysis, like a snapshot from a brain scan of a thinking person.

As Anil Seth proposes that we live in a "fleeting immersive inner movie (subjective experience), where we actively project reality by our best guess of what we observe and experience"

 

[someguy1]

But you won't know what you're doing, as far as transfer of the thinking ability.

No thinking ability exists when we take a snapshot of a computer's memory and transfer it to another computer and resume the computation.

I hope I didn't give the impression that I think the mind is a computer. On the contrary, I disagree with the claim that the mind is any kind of computation.

A neural net executing on a computer is a program like any other. If you freeze the cpu, dump RAM, load the dump on a similar computer (assuming the contents of the database etc. are also copied over) and then start up the computation, it will continue as if it were on the original machine.

I said NOTHING about minds. I hope this is clear.

The mystery of the "implementation" of mind will remain, only now it's a mystery of what a computer is doing.

No. There is never any mystery as to what a computer is doing, even one executing a complex program like a neural network. I agree that the mystery of mind remains. But mind is not a computation.

You can in principle duplicate the hardware and snapshot software status of any brain, but that doesn't tell you what you need to know.

You can do no such thing. For the reason that the brain is not a digital computer. For one thing, isn't it sort of a coincidence that the very technology that rules our world is now our explanation for mind? The Romans had great waterworks and thought the soul was a flowing substance. From nous we get the philosophical word for mind as well as the root of the word pneumatic.

The idea that the mind is a digital computer is no different. It's just a conceit of our particular technological age.

When the Go computer that has reached 9-dan professional rank plays a game, its moves are analyzed as consequences of its thinking - back-engineered, in a sense - just as a human player's are. One could of course download a copy of its state, in principle, at the moment it made its decision, but that wouldn't tell you what you want to know. At best it would provide informative hints for analysis,

That's true. You'd have to take the source code and a copy of the memory dump and a pencil and paper and analyze the hell out of it to figure out what the computer is going to do. That's also true if you took a memory dump of a 1960's mainframe running a payroll program. You couldn't figure out what check it's going to print next without the same type of painstaking analysis. In both cases it's perfectly deterministic, but super complicated.

like a snapshot from a brain scan of a thinking person.

No not at all. We don't know how the brain works. At best we can show you a picture of a cat and watch some area of your cortex light up That's a primitive level of understanding. We have no idea how the brain works and no proof or evidence, merely a theoretical speculation, that it's any kind of computation.

I hope my position is more clear. Computer programs can be easily transferred from one computer to another via memory dump. And even the most sophisticated neural net is a computer program.

On the other hand I deny that the mind is a program executing in the brain. And even if it is, we have no idea what the program is and we're nowhere close to having any kind of idea of how to determine that. Those saying the mind is a computer might as well say that the soul is like the great aqueducts of Rome. It's the same kind of culturally shortsighted belief.

 

[Write4U]

The idea that the mind is a digital computer is no different. It's just a conceit of our particular technological age

I disagree completely. Computational functions are a real phenomenon and seem to occur throughout the universe in both sentient and non sentient objects. In fact this mathematical function is what allowed our mathematics of scientific inquiries to be developed.

To assign an electro-chemical computational ability to the brain function is fundamental rather than conceit.

IMO, it is any other metaphysical interpretation of mind functions, that would speak of hubris. The human brain functions the same as all other sentient life, it is just the latest and most advanced model of computing skills. Our program does not just function on an "if/then" computation, but also includes a "why?" search command. Apparently this extraordinary ability is also present in other highly intelligent species, so IMO, this suggests that neural networks (synapses) are in an evolutionary state, where the greater connectivity leads to greater understanding of our reality.

Consider that Philosophy (abstract thinking) is founded on pure logic, and what is more logical than a form of computing system? Reality consists of observable patterns. Pattern recognition is a mathematical (in the broadest sense) computational function, be it from binary electrical coding (as in computers), or from an electro-chemical coding (as in organic brains).

Your brain is an enclosed biological computer, over which we have only part conscious control. All input the brain receives is an electro-chemical translation (coding) of sensory experiences, which it must combine to form an illusionary pattern (a best guess), which is then projected on the observed item. If the projection matches the object's properties with the brain's inner holograph, it confirms the match and we "know" what we are looking at.

I liked Anil Seth demonstration of some 50 representations of the letter R, in all its patterns by which this letter canbe symbolized. Yet one quick glance will tell us that each representation is the letter "R". This is a computational function based on prior experience with seeing the letter R in all its configurations. Except perhaps in some far-Eastern countries which use a different alphabet. But note, a conventional computers can translate English into many languages and vice versa. But is obvious from translated languages (even on this forum) that the computer can only make it's "best guess" of what it is translating. Not so very different from the human biological brain function.

 

[someguy1]

I disagree completely.

I understand. Many people disagree with my point of view on this topic.

Computational functions are a real phenomenon and seem to occur throughout the universe in both sentient and non sentient objects.

I'm not really sure. When you say computational, I understand you to mean the technical definition as in a Turing machine (or rather a physical implementation of a bounded TM). Perhaps you can give me a specific example of an undeniably computational (ie algorithmic) naturally occurring process.

In fact this mathematical function is what allowed our mathematics of scientific inquiries to be developed.

Mathematical functions allowed mathematics to be developed? I'm not sure I take your meaning.

To assign an electro-chemical computational ability to the brain function is fundamental rather than conceit.

Electrochemical, yes. Computational? That's a thesis but it awaits proof. If someone shows up with the actual algorithm and says: "Behold, this specific algorithm implements mind," and supplies PROOF, then I'll believe it. Till then, it's just a belief without evidence.

If I tell you that the Euclidean algorithm is an algorithm, I can show you the code. Euclid did so 2300 years ago. If you tell me the mind is an algorithm, then you CAN'T show me the code. That's a weakness in your argument.

IMO, it is any other metaphysical interpretation of mind functions, that would speak of hubris. The human brain functions the same as all other sentient life, it is just the latest and most advanced model of computing skills.

Opinion, not fact. Common opinion these days, to be sure. But you haven't got the algorithm so you are on weak ground in terms of actual evidence.

Our program does not just function on an "if/then" computation, but also includes a "why?" search command.

Can you show me the code? A Why command? You are just making this up. But you know you can ask a database management system to explain how it chose to execute a particular query. And you can program a neural net to explain why it made a particular move in a game. It's still just a computer program. If this then that.

Apparently this extraordinary ability is also present in other highly intelligent species, so IMO, this suggests that neural networks (synapses) are in an evolutionary state, where the greater connectivity leads to greater understanding of our reality.

Well neural nets are a model of computation and can be reduced to Turing machines, and you haven't got the code or any evidence for your claim.

Consider that Philosophy (abstract thinking) is founded on pure logic, and what is more logical than a form of computing system? Reality consists of observable patterns. Pattern recognition is a mathematical (in the broadest sense) computational function, be it from binary electrical coding (as in computers), or from an electro-chemical coding (as in organic brains).

Many argumentation fallacies in that para. Reality consists of observable patterns? Like constellations in the stars? Pattern recognition is mathematical therefore so is the universe? Well chess is played by humans and chess can be implemented as a computer program but it doesn't follow that humans can be so implemented. You're not giving a logical argument here, just a rhetorical one that I'm not buying.

Your brain is an enclosed biological computer,

Repetition of your thesis does not constitute an argument, let alone proof.

over which we have only part conscious control.

Freud made that point too but I don't think he was a computationalist.

All input the brain receives is an electro-chemical translation (coding) of sensory experiences, which it must combine to form an illusionary pattern (a best guess), which is then projected on the observed item. If the projection matches the object's properties with the brain's inner holograph, it confirms the match and we "know" what we are looking at.

Numerous studies show that that is NOT how the brain works. There's no "cat" picture in your brain that you compare to sensory impressions of a cat. It does NOT work that way. The brain's inner holograph? That's a false belief. There is no such thing. There are no pictures of cats in your brain, holographic or otherwise.

I liked Anil Seth demonstration of some 50 representations of the letter R, in all its patterns by which this letter canbe symbolized. Yet one quick glance will tell us that each representation is the letter "R". This is a computational function based on prior experience with seeing the letter R in all its configurations. Except perhaps in some far-Eastern countries which use a different alphabet.

Ok. But that doesn't form any part of an argument for your thesis.

But note, a conventional computers can translate English into many languages and vice versa.

As it happens I resided for a time out of the US, and I was a daily user of Google translate. I would always joke to myself that if this is the state of the art of AI, then I'm not worried.

But now we're back to the Chinese room. The room can translate English to Chinese and back again (far better in theory than Google translate can in practice) but the room does not understand Chinese or English. Of course people have been debating the Chinese room for decades so again we'll have to agree to disagree rather than litigate this here.

But is obvious from translated languages (even on this forum) that the computer can only make it's "best guess" of what it is translating.

Now you are starting to get it! Computers are actually terrible at translation. Not bad on a phrase or two, but awful when it comes to a nuanced article in a newspaper, for example.

Not so very different from the human biological brain function.

Incredibly different. If you don't think native understanding of a language is qualitatively different than machine translation, nothing I can say can convince you. We'll have to agree to disagree.

 

[Write4U]

Well chess is played by humans and chess can be implemented as a computer program but it doesn't follow that humans can be so implemented.

Chess is a mathematical game and if humans can play chess, then their mind is making "calculations".

Your insistence that requires a specific algorithm is mere speculation. At a pure electronic level, perhaps, but the human brain is more flexible than that. It creates algorithms on the fly. IOW, it can program itself.

question; is logic an algorithm?

And even then brain functions are "best guesses" also, just like a computer translation algorithms, which are fixed programs.

If a Lemur can tell the difference between "more of this" and "less of that" does it employ an algorithm?

p.s. your assigned the wrong poster's name to your quotes, although I would look forward to having James R respond also. I respect his opinion.

 

[billvon]

ANNs have a lot of popular mysticism about them. In fact they are entirely conventional computer programs, implemented on conventional hardware.

More and more, that's not true. Since so many applications are taking advantage of neural network topologies, purpose-built hardware is being included in chipsets specifically to run neural network algorithms. Often these are topologically very different from the classic Harvard or von Neumann architectures. For example, one of our processors has an NPE (neural processing engine) that runs such algorithms.

 

[Write4U]

This may be of interest.

How a Color-Blind Artist Became the World’s First Cyborg.
Neil Harbisson can "see" in UV thanks to an antenna-like implant that boosts his perception of light and gives him super-senses.

https://news.nationalgeographic.com...80108&utm_campaign=Content&utm_rd=24308569435

Can't do that with a computer, unless you write a translating program. In this case the brain itself made the adjustment.

 

[James R]

....

To to be clear, in the linked post someguy is reponding to what Write4U wrote, not to something I wrote. Quotes attributed to me in the linked post are not mine.

 

[Write4U]

To to be clear, in the linked post someguy is reponding to what Write4U wrote, not to something I wrote. Quotes attributed to me in the linked post are not mine.

I already brought that to his attention, but I would look forward to your impressions in regard to both posts.

 

[someguy1]

Chess is a mathematical game and if humans can play chess, then their mind is making "calculations".

Sure, just as a programmer stepping through the code with pencil and paper is emulating a computer. But not EVERYTHING humans do is explainable as an algorithm.

Your insistence that requires a specific algorithm is mere speculation. At a pure electronic level, perhaps, but the human brain is more flexible than that. It creates algorithms on the fly. IOW, it can program itself.

That's really a stretch. Barely merits a response. How does a computer program itself? Of course we have 4GL's such as database languages that specify the "what" and the program figures out the "how." But the idea of a program that programs itself on the fly to solve problems it's never seen before is just silly. The burden is on you to say something sufficiently rational as to merit a response in kind.

question; is logic an algorithm?

First-order logic can certainly be implemented as an algorithm.

And even then brain functions are "best guesses" also, just like a computer translation algorithms, which are fixed programs.

A reference to neurological literature would give you some credibility. You're just making stuff up now. I don't mean to come off as hostile but you are writing nonsense.

If a Lemur can tell the difference between "more of this" and "less of that" does it employ an algorithm?

I don't know anything about lemurs. I don't understand your point. An algorithm as commonly understood as the execution of a digital computer? No, I don't think so.

p.s. your assigned the wrong poster's name to your quotes, although I would look forward to having James R respond also. I respect his opinion.

My apologies to both of you for screwing up the quotes.

 

[someguy1]

To to be clear, in the linked post someguy is reponding to what Write4U wrote, not to something I wrote. Quotes attributed to me in the linked post are not mine.

My apologies.