Uncommon Descent Serving The Intelligent Design Community

CSI Confusion: Who Can Generate CSI?

Share
Facebook
Twitter
LinkedIn
Flipboard
Print
Email

In my first post, I discussed the importance of mechanism. In order to compute CSI you have to take into account the mechanism. Computing CSI without a mechanism is wrong. I deliberately focused on the use of specified complexity in evaluating various possible mechanisms. This is how Dembski uses CSI in his Design Inference argument. However, we are often interested in a system: a collection of artefacts and the mechanisms that operate on those artefacts. This is the context in which Dembski argues for the Law of Conservation of Information. Many of the questions that have come up are related to the context systems and who or what can generate CSI.

With a large probability, closed systems do not exhibit large increases in CSI. Small increases in CSI can be explained by “luck,” but large increases in CSI are too improbable to explain by recourse to simple “luck.” This means that any large increase in the CSI of a system must be explained by something external to that system.

Consider a wireless printer sitting in a room. If the printer begins to print out sonnets does the CSI of the room increase? The printer itself does nothing substantial to increase the probability of sonnets appearing. None of the mechanisms at work in the room explain sonnets. Thus the probability of sonnets appearing is very low and constitutes an increase in CSI. That’s because the explanation of the sonnets is external to the room itself.

Replace the printer with a sophisticated robot that sits in a room composing sonnets. As the robot produces sonnet after sonnet, does the CSI in the room increase? Given that the room contains a robot which is capable of composing sonnets, the production of sonnets is highly probable. The robot with the sonnets has approximately the same probability as the robot by itself. The improbability of the state of the room has not significantly increased due to the robot’s action. As such, the CSI of the room has not increased.

At this point, Sal objects that this is inelegant. The entropy of a system is the sum of the entropy of its parts. However, the CSI of a system is not the sum of the CSI of its parts. That’s an unavoidable consequence of using probabilities. They don’t combine in a conserving manner. It would be convenient and elegant if the CSI was the sum of the CSI of its parts. However, they don’t.

However, what if the sophisticated robot were replaced with a human poet? If the human poet composes sonnets, does the CSI in the room increase? The question comes down to whether or not the the human is part of the system. If the human is internal to the system, the probability of the sonnets has to be calculated taking into account his presence and thus the production of the sonnets will not be an improbable event. On the other hand, if the human is not internal to the system, then the CSI of the room is increasing. Everything hinges on whether or not the human is part of the system.

Dembski defines intelligence as the complement of chance and necessity. This means that intelligence cannot be reduced to random chance or natural law. It cannot be represented as a stochastic process. Put another way, we cannot put probabilities on the actions of intelligent agents. We might be able to do so in certain cases, but in general intelligent agents do not follow probabilities. As a consequence, an intelligent agent can never be internal to the system. In order to calculate CSI, we need to calculate probabilities and if there is an intelligent agent in the system, the probabilities will not exist. Thus, intelligent agents are always external to the system, and thus have the capability of inserting CSI into the system.

When applying CSI to a system, it tells us whether or not the system needs something external to the system in order to explain it. After defining the boundaries of the system, we can ask whether or not the system itself explains its own state. If it doesn’t, we know something external to the system must have influenced the system. Intelligence is taken to be external to the system because intelligence is difficult or impossible to reduce to probabilities. That its why we say intelligent agents generate CSI, but robots don’t.

Comments
scordova #15
a bacterial colony has a 100% probability of growing from 1 bacterium, thus the bits of information increase are 0 using Winston’s procedure. But that entirely misses the evidence of design that we have a copy machine as evidenced by the increase in Shannon information in the system and the numerous copies. Duplicates are strong evidence of design.
I explain how I see the issue. A single cell, say, the unique ancestor of the colony, has a huge complexity. Here I think we IDers all agree. You can use whatever complexity measure and you will find high values of complexity. In fact such cell is a self-copier automaton. Here - as Jerry said - "there is no alienation possible" for IDers, because, whatever measure we use, they all point to design. So, be quiet, here IDers are always right, compared to evolutionists who believe that cell arose by chance. Indeed because the cell design is to copy there is a 100% probability of growing a cell colony. Like, say, my new pen has a 100% probability of writing. Here you seem scandalized because "the bits of information increase are 0 using Winston’s procedure and that entirely misses the evidence of design that we have a copy machine". No. Dembskian CSI doesn't say at all that a self-copier automaton contains 0 bits. Differently, Dembskian CSI says that the output of a self-copier automaton, when frameworked in the context, add no new bits. In fact we have simply another copy of the cell. So there is no "missing the evidence of design that we have a copy machine". Here you can use Dembsky or algorithmic information or whatever and you will always find that the first self-reproducing automaton has huge information and the following copies add nothing. You say: "Duplicates are strong evidence of design". Yes, but duplicates are design evidence not per se, rather because manifest the power of the duplicator. The power of the duplicator is well accounted by any complexity measure. You seem to want that duplicates matter. But duplicates are only quantity and complexity measures somehow have to deal with quality, then complexity measures count zero the duplicates. In your post on the 2000 coins I already said that your 2000 coin heads count only 1 bit. To sum up it seems to me here you are causing "a storm in a glass of water".niwrad
November 26, 2013
November
11
Nov
26
26
2013
12:41 AM
12
12
41
AM
PDT
The incredibly low probability is obvious and has never been indicated otherwise.
Obviousness is a common but unpersuasive argument often offered around these parts.
Any 10 year old can understand the implications of it.
Run out and find me a ten year old child. I can't make head nor tail of it! /G Marks
Unfortunately because it is difficult to estimate exactly the incredible improbability it is discounted as meaningless. Sorry it is a prima facie indication of an incredibly low probabilistic event. Let the technicians calculate just how low it is.
Indeed! Let's have some kind of meaningful calculation. It's not as if requests for meaningful demonstrations of how to calculate CSI of, well, anything have not been made before.Alan Fox
November 25, 2013
November
11
Nov
25
25
2013
03:10 PM
3
03
10
PM
PDT
So, your pointing to conceptual notions involving infinity (as PI or e), seems a weak reason to adopt a new algorithmic metric, instead of the usual old Dembskian CSI metric or in addition to it.
That's not the reason to adopt a "new" metric. Also, it is not a "new" metric, but it is implicitly standard in the Information Science, maybe by other names and terms but the delineation exists between Algrorithmic Information and Shannon Information. ID literature and discussions on the net in particular try to conflate the two. As far as large numbers of digits, how about large bacterial colonies with trillions of cells. The Shannon information is tremendously larger than for one cell of bacteria, even though there is no net increase in algorithmic information of any complexity. Separating the Algorithmic from the Shannon is quite natural just as we separate the Shannon information in compressed files from the uncompressed forms even though the algorithmic information is identical (conserved). Instead of the robot example we could say, a bacterial colony has a 100% probability of growing from 1 bacterium, thus the bits of information increase are 0 using Winston's procedure. But that entirely misses the evidence of design that we have a copy machine as evidenced by the increase in Shannon information in the system and the numerous copies. Duplicates are strong evidence of design. In fact Bill Dembski's Design Inference references copyright infringement as an evidence of design, ergo copies of something can evidence design in the right context. Saying the increase was zero bits ignores the evidence of design of a built-in copier in the bacteria. Alternatively, and more sensibly, we can say the algorithmic information increases by essentially zero bits because it is a duplicate, there is no need to add the inelegance of mechanism and associated probabilities into the discussions. It adds confusion factors. And then we can say the Shannon information increases, and it looks like highly improbable duplicates (relative to random pre-biotic soups) arose, and thus design is signaled, suggesting the existence of a designed copy machine (the bacteria's self-replicating automaton). I will almost never agree with R0bb, but he says exactly what I might have said:
When IDists talk about the CSI in something, they are not referring to the CSI based on the actual mechanism in operation. If they were, then it would make no sense to talk about the CSI in designed artifacts, which IDists talk about all the time.
scordova
November 25, 2013
November
11
Nov
25
25
2013
12:54 PM
12
12
54
PM
PDT
gpuccio @8:
I don’t agree. As should be clear by now, I do believe that consciousness is absolutely necessary to generate new, original dFSCI. It is not a case that no computer algorithm can write original language, with new original meanings. The intuition of meaning is necessary for that. A machine, however sophisticated, has no idea of what meaning is. It also has no purposes. Algorithmic recombination of what already exists, without a perception of what it means, cannot generate new original complex meanings. I don’t believe that human beings create language and software and machine by simply using computing algorithms. I believe that all evidence we can observe supports the fundamental role of consciousness as the sole originator of dFSCI.
I perfectly understand your objection and you are absolutely correct in stating that no behaving entity can properly use language without having an understanding of the meaning of the words and phrases. Although I think you used the word 'meaning' without offering a definition, your objection is still valid from my perspective. This is something that the AI community is acutely aware of and they have a name for it: "symbol grounding". It simply means that symbols must represent (be associated with) perceptual experiences. There is no question that symbol grounding is beyond the capability of current intelligent machines. This is why IBM's Watson, although impressive, will often make simple mistakes that a human child would never make. But this does not mean that a computer program cannot learn to recognize patterns and associate the patterns with various symbols. Understanding meaning is not some metaphysical ability that lies in an unreachable realm. For the last few years, I have been working on a program called Rebel Speech that can learn to recognize various sounds such as dog barks and car horns (in addition to speech) all by itself after being exposed to the sounds of television programs. Sure, it cannot associate those sounds with symbols (I had to do it manually) but the ability to automatically associate symbols with perceptual patterns is well within the realm of possibility. It's a temporal correlation, i.e., it's all in the timing. Even the ability to recognize metaphors and analogies, which is essential to language, is based on temporal correlations and can be formalized and automated by a mechanism.
Burt, of course, you are entitled to your own opinion, and I am perfectly aware that many would not agree with me.
I do indeed disagree with your position on this issue.Mapou
November 25, 2013
November
11
Nov
25
25
2013
11:17 AM
11
11
17
AM
PDT
Sal and niwrad: Again, I don't see the problem.Let's take again pi as an example. The concept of pi is obviously the specification. It has neither bits nor complexity. It is a conscious concept, a definition of a conscious intelligent agent. Any coded sequence which corresponds to n digits of pi has a complexity which depends on its length, and is obviously specified by the correspondence to the concept of pi. It is equal to the specific result of a computation based on the concept. An algorithm which can compute n digits of pi has a complexity which depends on the minimun number of bits necessary to accomplish that function. In any system where digital sequences can arise by random variation, the complexity of a given sequence of n digits of pi can be higher or lower than the complexity of the algorithm which can compute it. If it is lower, that is the dFSCI of that sequence. If it is higher, then the complexity if the algorithm is indeed the Kolmogorov complexity of the sequence, because the algorithm can be considered a compression of the sequence (once the algorithm exists, the sequence will be necessarily generated. So, again, I can't see where is the problem. We must never mix up the specification (the concept of pi) with the complexity of a given digital sequence. And we must distinguish between simple complexity and Kolmogorov complexity. Niwrad, I agree with you that with biological sequences, like protein sequences, all those reasonings are however rather useless. Biological sequences are random-like, scarcely compressible. The only algorithm which can generate them is a very complex algorithm which already knows the function to be generated, and which can compute the right sequence by a perfect knowledge of biochemical laws. Or by random variation + intelligent selection (measurement of the desired function, and gradual approximation to it by selection). But both thses kinds of algorithms are much more complex than the sequence itself, and certainly they are not naturally available in a biological context. There is, however, a good example of a very complex biological algorithm, embedded in living beings, which can refine proteins by random variation plus intelligent selection based on information derived from the environment: it is the well known example of antibody maturation after the first immune response. But that is another story.gpuccio
November 25, 2013
November
11
Nov
25
25
2013
11:11 AM
11
11
11
AM
PDT
addressing these issues at the risk of alienating some of their fellow IDists.
There is no alienation possible. The incredibly low probability is obvious and has never been indicated otherwise. Just estimating exactly what the probability is may be an issue but not that it is so near zero that the main issue is estimating the number of zeros after the decimal place.jerry
November 25, 2013
November
11
Nov
25
25
2013
11:04 AM
11
11
04
AM
PDT
Unfortunately Jerry is also correct here:
Now FCSI or however kairosfocus labels it is easy to understand.
because GEM’s calculation is trivial and useless as an exercise, merely a log transformation of a number simply based on the count of residues in a protein sequence. It is far from useless and an example of immense complexity that specifies another function. Any 10 year old can understand the implications of it. Unfortunately because it is difficult to estimate exactly the incredible improbability it is discounted as meaningless. Sorry it is a prima facie indication of an incredibly low probabilistic event. Let the technicians calculate just how low it is.jerry
November 25, 2013
November
11
Nov
25
25
2013
11:00 AM
11
11
00
AM
PDT
scordova #6
I’m suggesting if CSI is measured in Shannon metrics it’s probably not the best way to measure conceptual notions. What are conceptual notions. The number PI or the number e, these are conceptual notions, their information content is invariant, but we really don’t speak in terms of measuring the exact bits in PI or e, only in terms of the most compact algorithm.
Your examples, PI or e, are numbers with an infinite number of different digits (not repeated as in your example of 2000 coins!). As such their CSI, measured in Shannon metrics, should be infinite. For this reason you are right that, for them, it would be more appropriate the algorithmic metric. But ID theory (in its application to biology) is not much interested in measuring the CSI of conceptual mathematical notions, rather of finite sequences and, eventually, finite systems. So, your pointing to conceptual notions involving infinity (as PI or e), seems a weak reason to adopt a new algorithmic metric, instead of the usual old Dembskian CSI metric or in addition to it.niwrad
November 25, 2013
November
11
Nov
25
25
2013
10:43 AM
10
10
43
AM
PDT
scordova:
I thank Winston for raising this issue. It was a courageous move since it was sure to spark public disagreement in the UD community. I knew I would spark disagreement with my last series of posts that began the current discussion.
I agree. Sal and Winston are to be commended for addressing these issues at the risk of alienating some of their fellow IDists.R0bb
November 25, 2013
November
11
Nov
25
25
2013
10:13 AM
10
10
13
AM
PDT
Mapou: I don't agree. As should be clear by now, I do believe that consciousness is absolutely necessary to generate new, original dFSCI. It is not a case that no computer algorithm can write original language, with new original meanings. The intuition of meaning is necessary for that. A machine, however sophisticated, has no idea of what meaning is. It also has no purposes. Algorithmic recombination of what already exists, without a perception of what it means, cannot generate new original complex meanings. I don't believe that human beings create language and software and machine by simply using computing algorithms. I believe that all evidence we can observe supports the fundamental role of consciousness as the sole originator of dFSCI. Burt, of course, you are entitled to your own opinion, and I am perfectly aware that many would not agree with me.gpuccio
November 25, 2013
November
11
Nov
25
25
2013
09:48 AM
9
09
48
AM
PDT
i.e. you can increase the algorithmic information in the concept of PI, it is invariant.
ERROR: you can not increase the algorithmic information in the concept of PI, it is invariant. SORRY!scordova
November 25, 2013
November
11
Nov
25
25
2013
09:41 AM
9
09
41
AM
PDT
I thank Winston for raising this issue. It was a courageous move since it was sure to spark public disagreement in the UD community. I knew I would spark disagreement with my last series of posts that began the current discussion. The argument is not who is right or wrong, but which conventions are most workable in trying to conceptualize the evolution of information. I'm suggesting if CSI is measured in Shannon metrics it's probably not the best way to measure conceptual notions. What are conceptual notions. The number PI or the number e, these are conceptual notions, their information content is invariant, but we really don't speak in terms of measuring the exact bits in PI or e, only in terms of the most compact algorithm.
At this point, Sal objects that this is inelegant. The entropy of a system is the sum of the entropy of its parts. However, the CSI of a system is not the sum of the CSI of its parts. That’s an unavoidable consequence of using probabilities. They don’t combine in a conserving manner. It would be convenient and elegant if the CSI was the sum of the CSI of its parts. However, they don’t.
The resolution is to separate Information in evidence in an artifact (measure it in Shannon entropy) and then some offer some estimate of the information of the concept (for example PI can be represented with N bits in a particular language). We can then say the Algorithmic information (which is K-complex) is conserved but the information levels of artifacts can increase. For example with the robot, the concept of "all heads" is inside the robot, that is invariant just like PI. Whereas when he orders all the coins, the information in the coins increases. Trying to force fit Shannon entropy to the entire system is like one size fits all. You can force it to work, but it's not elegant, it's not the way the general public thinks, nor even a lot of IT and physicists. I'm not suggesting a radical re-write of ID literature, this is minor adjustment that will relieve a lot of indigestion. Now with respect to AI and real intelligence: As far as AI that is implemented on Turing Machines, they have finite states even though they can go into infinite loops or not halt, they are still finite state machines. A program running an infinite loop can still be described with a finite code. For example we can have an infinite loop program pumping out the digits of PI forever (or until the plug is pulled). How do we contrast this to a real intelligence. Let me suggest we liken the AI system as a fixed-limited energy source, and the real intelligence (or any comparable black box) like an open channel of new energy whose bound is not known. The AI system has a fixed-limited amount of conserved information it can put in the system (Algorithmic Information). It can increase the Shannon information in a system, but not increase the algorithmic information in the system. i.e. you can increase the algorithmic information in the concept of PI, it is invariant. A real intelligence, which may have access to unforeseen insights and new ideas can be modeled as an open channel of new algorithmic information. Ideas after all cannot be necessarily measured in terms of exact bits (like the number PI cannot be measured in terms of exact bits, but the concept is quite real in the mathematical sense even though we can't exactly count its information content). This sketches out what I would suggest are workable conventions and conventions that can be taught in the classroom. Jerry echoes exactly the consternation that has been simmering below the surface over CSI in the ID community. I'd say the root problem is making CSI the one-size fits all for all information. I'm complaining that even though some can force such a convention to work, maybe another convention might be more workable. I'm saying we need two sizes instead of one: 1. one for algorithmic information 2. one for Shannon information in artifacts The robot example, the decompression of ZIP files, the outputting of PI, bacteria multiplying into a colony -- these can be more clearly conceptualized if we have approach to information with two conceptions. This is not outrageous given we can have a compressed MP3 file that is 1 Gig and its decompressed form is 30 Gigs. The convention I'm suggesting resolves the differing metrics of information for each of the files even though the algorithmic information for each of the files is identical. It also gives a way to deal with concepts with invariant algorithmic information content like PI, e, or other abstract ideas.scordova
November 25, 2013
November
11
Nov
25
25
2013
09:39 AM
9
09
39
AM
PDT
Winston, thanks for spelling out the issues clearly, but there still seems to be some confusion. In your previous post, you addressed the robot-in-a-room scenario as follows:
For that matter, take a robot in a room full of coins that have random heads-tail configurations. The robot orders them all heads. The final CSI inside the room (the Robot’s CSI plus the coin’s CSI) is now greater than what we began with!
Remember that the CSI has to calculated based on the actual mechanism in operation. In this case, we have to calculate the CSI taking into account the actions of the robot.
But of course it's incorrect to say that CSI has to be calculated based on the actual mechanism in operation. Rather, a CSI calculation is based on a hypothesized mechanism, usually with the hope that the calculation will tell us that the hypothesized mechanism was not the actual mechanism in operation. When IDists talk about the CSI in something, they are not referring to the CSI based on the actual mechanism in operation. If they were, then it would make no sense to talk about the CSI in designed artifacts, which IDists talk about all the time. This confusion carries over to the current post, with the sonnet-writing robot:
As the robot produces sonnet after sonnet, does the CSI in the room increase? Given that the room contains a robot which is capable of composing sonnets, the production of sonnets is highly probable. The robot with the sonnets has approximately the same probability as the robot by itself. The improbability of the state of the room has not significantly increased due to the robot’s action. As such, the CSI of the room has not increased.
Again you're assuming that the CSI is calculated based on the actual mechanism in operation. But that is not what IDists mean when they talk about the amount of CSI in something.R0bb
November 25, 2013
November
11
Nov
25
25
2013
09:28 AM
9
09
28
AM
PDT
In order to calculate CSI, we need to calculate probabilities and if there is an intelligent agent in the system, the probabilities will not exist. Thus, intelligent agents are always external to the system, and thus have the capability of inserting CSI into the system.
Does Winston Ewert mean to say he can produce a meaningful CSI calculation? Maybe even give us an example of a calculation applied to some biological entity or system? Jerry is quite correct when he says:
People have been discussing CSI on this site for several years and the consensus is that no one can explain just what it is. I believe that is still true.
Unfortunately Jerry is also correct here:
Now FCSI or however kairosfocus labels it is easy to understand.
because GEM's calculation is trivial and useless as an exercise, merely a log transformation of a number simply based on the count of residues in a protein sequence.Alan Fox
November 25, 2013
November
11
Nov
25
25
2013
09:21 AM
9
09
21
AM
PDT
People have been discussing CSI on this site for several years and the consensus is that no one can explain just what it is. I believe that is still true. This post confuses as opposed to enlightens. Maybe there is a simple explanation somewhere. I just have not seen it. Now FCSI or however kairosfocus labels it is easy to understand.jerry
November 25, 2013
November
11
Nov
25
25
2013
08:59 AM
8
08
59
AM
PDT
gpuccio:
My point is: a robot, however sophisticated, can never compose original sonnets, because a robot is not conscious and cannot understand meaning. It can, certainly, mix and recycle existing phrases according to pre-inputted rules, but in that case all the complexity has already been given to it by intelligent agents, a no new original CSI is created. The process is algorithmic, ar a mix of algorithmic and random, but it cannot generate new original complex meaning. IOWs, I simply don’t believe that any algorithmic, non conscious process can generate new original CSI, beyond what is already in the system. Obviously, some specified information can be randomly generated, but it will never be “complex”, if we use an appropriate threshold for complexity.
Interesting. In my opinion, if the robot combines a random selection process with either pre-programmed or learned rules, there is no question that new CSI can be created and there is no reason to suppose that the information thus created cannot be complex. I am prepared to argue that a human poet uses the same method: randomly combining words according to rules of grammar, semantics and sonnet-specific rules. It's a simple matter of asking what-if questions. The only difference is that a human poet can recognize beauty in the newly created sonnets whereas the robot cannot. The reason is that beauty is not a property of matter that can be formalized although a robot can be taught to recognize a type of beauty. Beauty is a spiritual property.Mapou
November 25, 2013
November
11
Nov
25
25
2013
08:53 AM
8
08
53
AM
PDT
Winston Ewert: My views are slightly different. You say:
Consider a wireless printer sitting in a room. If the printer begins to print out sonnets does the CSI of the room increase? The printer itself does nothing substantial to increase the probability of sonnets appearing. None of the mechanisms at work in the room explain sonnets.
Indeed, for the printer to print the sonnets, it must receive the information. If it is wireless, it can receive the information from outside, but that is rather irrelevant. The CSI here is in the sonnets, not in the room. The relevant question is: does the specified information in the sonnets arise in the room? The answer is obviously not.
Replace the printer with a sophisticated robot that sits in a room composing sonnets. As the robot produces sonnet after sonnet, does the CSI in the room increase?
My point is: a robot, however sophisticated, can never compose original sonnets, because a robot is not conscious and cannot understand meaning. It can, certainly, mix and recycle existing phrases according to pre-inputted rules, but in that case all the complexity has already been given to it by intelligent agents, a no new original CSI is created. The process is algorithmic, ar a mix of algorithmic and random, but it cannot generate new original complex meaning. IOWs, I simply don't believe that any algorithmic, non conscious process can generate new original CSI, beyond what is already in the system. Obviously, some specified information can be randomly generated, but it will never be "complex", if we use an appropriate threshold for complexity.
However, what if the sophisticated robot were replaced with a human poet? If the human poet composes sonnets, does the CSI in the room increase?
I would definitely say yes. The sonnet simply did not exist before. Its information and meaning are generated by the conscious processes of the poet. CSI is generated, and it is absolutely irrelevant is we consider the room, or the world, or the universe. When Shakespeare wrote his sonnets, CSI was generated in reality, because that specified form was the output of his conscious experiences. It did not exist before. It has existed since then. It is not important how many times they were printed, or where. The information came into existence in the moment Shakespeare conceived of it, and for the first time outputted that form to a sheet of paper. The rest is only algorithmic transfer.
Put another way, we cannot put probabilities on the actions of intelligent agents. We might be able to do so in certain cases, but in general intelligent agents do not follow probabilities.
That's exactly the point. The actions of intelligent agents come from their conscious representations, which defy mere causal or probabilistic description. Consciousness is transcendental. That's why a poet can write sonnets, and a robot cannot. That's why CSI, in all its forms, is always the output of conscious agents. That's why we can infer design from CSI. Indeed, the only possible definition of design is: the output of form from conscious representations to material objects.gpuccio
November 25, 2013
November
11
Nov
25
25
2013
08:32 AM
8
08
32
AM
PDT

Leave a Reply