Emergence

[durangodawood]

Just life seems to me like the classic case of emergence.

Whole new categories of rules that are totally novel compared those governing the 'inert' matter that life is built from.

 

[KCfromNC]

I just read an interesting essay on Connectionism at the Stanford Encyclopedia, which basically regards the conceptual issues surrounding neural networks. I didn't realize all of the challenges fuzzy logic is presenting to traditional computing.

Just to be clear, what you're discussing is debates among cog-sci or philosophy of mind types, not computer science people.

Neural nets are just linear algebra and some calculus - at least at the "how it might fundamentally change the basic theories of computation" level. Nothing particular challenging there that isn't seen in lots of other code.

 

[Resha Caner]

We can't predict the pattern with modern technology, but is it only a problem of measurement?

Possibly, but I don't think so. Or, it would be better to say I wouldn't be interested in emergence if that were the issue.

However, at some point, we may run into observer/measurement problems. In order to accurate measure the state of neurons we may end up causing significant changes that influence the outcome. Perhaps this is another property of emergence in itself.

Except maybe for this. This is an insightful observation.

 

[Resha Caner]

Just to be clear, what you're discussing is debates among cog-sci or philosophy of mind types, not computer science people.

Neural nets are just linear algebra and some calculus - at least at the "how it might fundamentally change the basic theories of computation" level. Nothing particular challenging there that isn't seen in lots of other code.

As long as the underlying hardware architecture doesn't change, that may be true. But if it drives new computing architectures, the landscape could change drastically. Quantum computing and self-assembly computing are two very interesting areas of research.

 

[KCfromNC]

As long as the underlying hardware architecture doesn't change, that may be true. But if it drives new computing architectures, the landscape could change drastically. Quantum computing and self-assembly computing are two very interesting areas of research.

They are, but they have nothing to do with neural nets. We already know what architecture is good for neural nets - they've been running in graphics processors for decades.

 

[Chris B]

Just life seems to me like the classic case of emergence.

Whole new categories of rules that are totally novel compared those governing the 'inert' matter that life is built from.

There is that.

 

[Resha Caner]

They are, but they have nothing to do with neural nets. We already know what architecture is good for neural nets - they've been running in graphics processors for decades.

I guess I don't rule out the possibility that it might spark a new idea at some point.

I think it understates the intricacy of nets to shrug them off as just some algebra and calculus. It occurred to me that their potential to bifurcate could make them emergent & unpredictable. I've always been intrigued by the way bifurcations are represented in code (which is supposed to be repeatable) versus the way they manifest in the physical systems that code is meant to represent.

 

[KCfromNC]

I guess I don't rule out the possibility that it might spark a new idea at some point.

Nothing's impossible, but there are certainly way more fruitful areas to explore - at least according to the people actually doing the work.

I think it understates the intricacy of nets to shrug them off as just some algebra and calculus. It occurred to me that their potential to bifurcate could make them emergent & unpredictable. I've always been intrigued by the way bifurcations are represented in code (which is supposed to be repeatable) versus the way they manifest in the physical systems that code is meant to represent.

I'm not sure how to politely put this, but huh?

 

[Resha Caner]

I'm not sure how to politely put this, but huh?

I didn't realize you worried about being polite. Since I don't know what you know or what you think is unintelligible about my statement, we'll have to tease it out. So, likewise, the question that follows is not meant to be insulting, but we have to establish common ground somehow.

Do you know what a bifurcation is?

 

[KCfromNC]

Do you know what a bifurcation is?

Yes. But all software has conditional code, and combined with the regular structure of NN code and relative lack of data dependent branches I'm surprised you'd think that would be difficult for modern branch prediction hardware.

Or did you mean something other than the computer science version of the term?

 

[FrumiousBandersnatch]

That was the example that occurred to me, having tinkered in the past with Conway's game of life.

The initial rules are simple and do not suggest gliders, glider guns, organism arrangements to produce a series of glider guns...

Yes, I think cellular automata, particularly C's GOL, are excellent examples of emergence, because the active elements, the cells in the grid, are static, having only two states (on/off, alive/dead, black/white), and the rules specify whether a cell is on or off depending on its immediate neighbors. There's nothing apparently interesting in individual cells being on or off. It's the patterns of activity that are interesting; the patterns seem to have a 'life' of their own, moving across the grid and interacting in complex ways completely different from the static binary states of their components. The other notable feature is that the patterns that appear are dependent, not on the grid functionality per se, but on the initial on/off states of the grid cells - the 'seeds of life'. The emergent patterns are at a higher level of abstraction than the states of cells in the grid.

Flocking and shoaling behaviour are other good examples of emergence, which are also unpredictable from the individual creatures because the flocking or shoaling behaviours (rules) are only adopted when the individuals are in a shoal or flock.

 

[Resha Caner]

Or did you mean something other than the computer science version of the term?

I'm not an expert in computer science, but I meant something other than conditional code. I was thinking of the nonlinear nature of the activation functions, which can produce bifurcations like those described here:
https://en.wikipedia.org/wiki/Bifurcation_theory

A common example is the period doubling for the logistic map:
https://en.wikipedia.org/wiki/Period-doubling_bifurcation

The example still has known equilibria based on specific values of x and r. In other words, a computer will produce the same result every time.

However, there are models of physical systems (such as the Duffing oscillator) where the mathematics yield bifurcations. In some cases it even implies multiple possible equilbria for the same parameters. The question is whether the mathematics implies a real physical possibility? If so, how would the real physical system choose among the multiple equilbria? Or would it?

It seems nets that can bifurcate in this same way would have the same predictability issues.

 

[KCfromNC]

I'm not an expert in computer science, but I meant something other than conditional code. I was thinking of the nonlinear nature of the activation functions,
which can produce bifurcations like those described here:

Could be a reason why ReLU is a more common choice today.

However, there are models of physical systems (such as the Duffing oscillator) where the mathematics yield bifurcations. In some cases it even implies multiple possible equilbria for the same parameters. The question is whether the mathematics implies a real physical possibility? If so, how would the real physical system choose among the multiple equilbria? Or would it?

It seems nets that can bifurcate in this same way would have the same predictability issues.

Could be, but it sure seems the people working on them aren't all that keen to reject the math underlying them : http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6002509

 

[Resha Caner]

Could be a reason why ReLU is a more common choice today.

Could be. I haven't studied rectifiers with respect to whether they could possibly bifurcate.

I think there are other reasons as well.

Could be, but it sure seems the people working on them aren't all that keen to reject the math underlying them

I wasn't suggesting the math should be rejected.

 

[Loudmouth]

Another example of emergence that is more cogent to this subforum are genetic algorithms. This is where random changes are thrown away or kept judged solely by function. Engineers have been able to construct more efficient circuits using this method. At the same time, those engineers are not always able to immediately figure out exactly how the circuit functions as it does. Here is one example:

A field-programmable gate array, or FPGA for short, is a special type of circuit board with an array of logic cells, each of which can act as any type of logic gate, connected by flexible interlinks which can connect cells. Both of these functions are controlled by software, so merely by loading a special program into the board, it can be altered on the fly to perform the functions of any one of a vast variety of hardware devices.

Dr. Adrian Thompson has exploited this device, in conjunction with the principles of evolution, to produce a prototype voice-recognition circuit that can distinguish between and respond to spoken commands using only 37 logic gates - a task that would have been considered impossible for any human engineer. He generated random bit strings of 0s and 1s and used them as configurations for the FPGA, selecting the fittest individuals from each generation, reproducing and randomly mutating them, swapping sections of their code and passing them on to another round of selection. His goal was to evolve a device that could at first discriminate between tones of different frequencies (1 and 10 kilohertz), then distinguish between the spoken words "go" and "stop".

This aim was achieved within 3000 generations, but the success was even greater than had been anticipated. The evolved system uses far fewer cells than anything a human engineer could have designed, and it does not even need the most critical component of human-built systems - a clock. How does it work? Thompson has no idea, though he has traced the input signal through a complex arrangement of feedback loops within the evolved circuit. In fact, out of the 37 logic gates the final product uses, five of them are not even connected to the rest of the circuit in any way - yet if their power supply is removed, the circuit stops working. It seems that evolution has exploited some subtle electromagnetic effect of these cells to come up with its solution, yet the exact workings of the complex and intricate evolved structure remain a mystery (Davidson 1997).
http://www.talkorigins.org/faqs/genalg/genalg.html#examples:electrical
​

 

[Resha Caner]

Another example of emergence that is more cogent to this subforum are genetic algorithms. This is where random changes are thrown away or kept judged solely by function. Engineers have been able to construct more efficient circuits using this method. At the same time, those engineers are not always able to immediately figure out exactly how the circuit functions as it does. Here is one example:

A field-programmable gate array, or FPGA for short, is a special type of circuit board with an array of logic cells, each of which can act as any type of logic gate, connected by flexible interlinks which can connect cells. Both of these functions are controlled by software, so merely by loading a special program into the board, it can be altered on the fly to perform the functions of any one of a vast variety of hardware devices.

Dr. Adrian Thompson has exploited this device, in conjunction with the principles of evolution, to produce a prototype voice-recognition circuit that can distinguish between and respond to spoken commands using only 37 logic gates - a task that would have been considered impossible for any human engineer. He generated random bit strings of 0s and 1s and used them as configurations for the FPGA, selecting the fittest individuals from each generation, reproducing and randomly mutating them, swapping sections of their code and passing them on to another round of selection. His goal was to evolve a device that could at first discriminate between tones of different frequencies (1 and 10 kilohertz), then distinguish between the spoken words "go" and "stop".

This aim was achieved within 3000 generations, but the success was even greater than had been anticipated. The evolved system uses far fewer cells than anything a human engineer could have designed, and it does not even need the most critical component of human-built systems - a clock. How does it work? Thompson has no idea, though he has traced the input signal through a complex arrangement of feedback loops within the evolved circuit. In fact, out of the 37 logic gates the final product uses, five of them are not even connected to the rest of the circuit in any way - yet if their power supply is removed, the circuit stops working. It seems that evolution has exploited some subtle electromagnetic effect of these cells to come up with its solution, yet the exact workings of the complex and intricate evolved structure remain a mystery (Davidson 1997).
http://www.talkorigins.org/faqs/genalg/genalg.html#examples:electrical
​

Cool ... though it still sounds conceptually like a net - just that a genetic algorithm was used to train it.

What I wonder is what happens when a sound other than "stop" or "go" hits the input side. If someone were to say "red" and "green" the net will make a decision. It may decide to stop for one of those words and go for the other. If there is a "no decision" option, Thompson hasn't created a system to recognize 2 words. He's created a system to categorize sounds into 2 (or 3) buckets.

The intriguing question would then be: Can a human recognize a pattern? i.e. would the system's categories make sense to us?

 

[Loudmouth]

Cool ... though it still sounds conceptually like a net - just that a genetic algorithm was used to train it.

I would say that it is very much like a neural net, although my training is more in biology than engineering.

What I wonder is what happens when a sound other than "stop" or "go" hits the input side.

I would strongly suspect that you would get mixed results. This examples is from 1997, so it is going to be primitive compared to what we have now. At the same time, "deep learning" and other applications of neural nets have been used by companies like Apple and Google to train their voice recognition algorithms. Wired.com has a ton of interesting articles on the current revolution of AI's in search and other sectors of programming. They are even building AI's that can build other AI's. As with the circuit in the example above, they often fail to understand how the AI written software works, which is again interesting for someone who is not a programming or electrical engineering expert.

One of my favorite articles is about an AI that they taught to play the game Go. AI's had beaten world class chess players in the past, but no AI had been able to beat world class Go players . . . until now. Some of the moves the AI made were completely counter-intuitive, but they were crucial for the win. I find that fascinating.

http://www.wired.com/2016/01/in-a-huge-breakthrough-googles-ai-beats-a-top-player-at-the-game-of-go/

If someone were to say "red" and "green" the net will make a decision. It made decide to stop for one of those words and go for the other. If there is a "no decision" option, Thompson hasn't created a system to recognize 2 words. He's created a system to categorize sounds into 2 (or 3) buckets.

I fully agree. Nonetheless, the GA solution worked better and worked with fewer components than the directly engineered systems of the same era. One of the emergent properties of GA's that have always intrigued me is that they find solutions that rational engineers would normally never think of using.

"How does it work? Thompson has no idea, though he has traced the input signal through a complex arrangement of feedback loops within the evolved circuit. In fact, out of the 37 logic gates the final product uses, five of them are not even connected to the rest of the circuit in any way - yet if their power supply is removed, the circuit stops working."
http://www.talkorigins.org/faqs/genalg/genalg.html#examples:electrical

The intriguing question would then be: Can a human recognize a pattern? i.e. would the system's 3 categories make sense to us?

Without training? I don't know. I am probably using the word too much, but language is also a fascinating topic when it comes to human intelligence. Children who are not raised around other humans (i.e. "feral" children) and do not learn language have measurable cognitive problems. It's as if the human brain requires the experience of language in order to develop properly. Language might be the very thing that programs the human brain in the same way that Google feeds thousands of pictures of kittens to its AI's in order to train them how to recognize pictures of kittens.

 

[Resha Caner]

Language might be the very thing that programs the human brain in the same way that Google feeds thousands of pictures of kittens to its AI's in order to train them how to recognize pictures of kittens.

Yes, nets use large sets to train, but as I understand it, that's not how people learn. Kids don't see thousands of kittens before they start identifying what is and isn't a kitten. From what I've read, people learn much more quickly if meaning is attached. Also, when memories are recalled, they don't seem to be recalled as literal impressions of an event (as they would be in a net). Rather, the "essence" of the event is recalled and the memory reconstructed based on the subject's current state.

I fully agree. Nonetheless, the GA solution worked better and worked with fewer components than the directly engineered systems of the same era. One of the emergent properties of GA's that have always intrigued me is that they find solutions that rational engineers would normally never think of using.

That is interesting. But isn't there a difference between "we wouldn't have thought of that" and "unpredictable"? It seems possible that with enough effort someone could reverse engineer some rules for your example and everyone would nod their heads and agree those are the principles the net is using to create its categories. But that feels like cheating.

An interesting follow-up would have been to change the architecture of the neurons in the net, repeat the exercise of training it to distinguish "stop" and "go", and then reverse engineer that second net to determine its principles. If it used the same principles as the first net, I would be willing to agree the emergence was predictable. If, however, it was using different principles, I would probably say the emergence wasn't predictable.

 

[Loudmouth]

Yes, nets use large sets to train, but as I understand it, that's not how people learn. Kids don't see thousands of kittens before they start identifying what is and isn't a kitten. From what I've read, people learn much more quickly if meaning is attached.

The analogy I was making is that the process of learning is what wires the brain.

Also, when memories are recalled, they don't seem to be recalled as literal impressions of an event (as they would be in a net). Rather, the "essence" of the event is recalled and the memory reconstructed based on the subject's current state.

Very much so. Memory isn't something that is like a VCR tape. It is more about associations and reconstructing events using those associations. It also involves emotions which can change our memory, either in the past or in the present.

That is interesting. But isn't there a difference between "we wouldn't have thought of that" and "unpredictable"?

I would say yes and no. In the end, what we label as unpredictable is based on human knowledge and intuition. As our knowledge changes and grows we are able to predict things that we weren't able to predict before. It is hard to differentiate between unpredictable and can't predict right now.

However, this is different than the question of how human engineers approach rational design. Unpredictable and counterintuitive are not necessarily the same thing.

An interesting follow-up would have been to change the architecture of the neurons in the net, repeat the exercise of training it to distinguish "stop" and "go", and then reverse engineer that second net to determine its principles. If it used the same principles as the first net, I would be willing to agree the emergence was predictable. If, however, it was using different principles, I would probably say the emergence wasn't predictable.

Given the strange unconnected parts that were vital to the function of the circuit, I would suspect that even changes in the individual gates themselves would change the outcome. A part with slightly different resistance or capacitance might change the outcome, for example. We start running into the butterfly effect, where even the smallest changes can produce wildly different outcomes.

 

[Resha Caner]

The analogy I was making is that the process of learning is what wires the brain.

OK.

Unpredictable and counterintuitive are not necessarily the same thing.

Agreed.

It is hard to differentiate between unpredictable and can't predict right now.

It is. I was going for a sense of unpredictable that is independent of us - of our knowledge. Such a thing would have to be established mathematically, but I don't know if it would be possible to achieve.

Maybe a different way of coming at it is the whole P vs. NP thing and the idea of reversible processes. For example, there isn't a general algorithm for factoring a number into primes ... at least not one other than guessing and verifying, and guessing isn't really a process. So maybe in that sense a net trained to factor numbers (which is something people try to do) is unpredictable.

... Even then P vs. NP hasn't been solved. Unless it were definitively shown that P /= NP, even that would be uncertain.