My Model of Human Consciousness

[rocketdodger]

Ah, so it turns out a Boltzmann Machine is basically a Hopfield Network that relies upon stochastic behavior to overcome being stuck in local minima. You should have just said that! Fascinating stuff ... I should have followed links to it sooner.

Yes, I am familiar with simulated annealing, lol. I was confused by the term "relaxation" although now that I think about it, it does apply well, since the system settles on energy minima. I wonder if thats why they label certain steps in minimum distance search algorithms as "relaxation" (the only place I had previously heard of it was in algorithms like Djikstra's or Floyd-Warshall or however you spell it) ...

So it turns out this is actually reinforces what I am trying to say here, although it does illustrate how badly I am communicating my point. I will explain in a post later today since I have to visit the mall with the in-laws ( ...) right now, but the gist of it is that the idea of associative memory was what I had in mind for how a "symbol" firing would lead to other "symbols" firing in a way that can be thought of as inference.

 

[mimada]

First of all, I would like to say that I applaud your efforts. You seem to have done a considerable amount of research in the formulation of your ideas. I would like to address an issue, however. In my opinion, you need to be more descriptive of your terminology. Even if you dismiss some criticism as being disingenuous, it seems others are misinterpreting your usage of particular words. For example, you are using "symbol" specifically to describe information compressed in a NN but CapelDodger has apparently interpreted it in the more conventional sense as a "object, picture, written word, sound, or particular mark that represents something else by association, resemblance, or convention." (from wikipedia). I know that you have entreatied everyone to read through your post completely before responding but it is the nature of humans to default to common usage definitions (this is perhaps an example of Bayesian networks in operation).

I do not want to see this discussion devolve into a philosophical debate about what conscious is so perhaps you could instead describe the properties inherent in conscious and how it pertains to your thesis. Also, it may be helpful to differentiate between human conscious and non-human conscious. I make mention of this because you made the distinction in your thesis and because a comparison may further clarify the concepts in discussion.

 

[INRM]

Malerin,

If you think that's unethical, how do you feel about destroying a conscious washing machine?

First of all no washing machine to the best of my knowledge is conscious or sentient.

However since this is obviously a hypothetical, I will say that you are looking at this the wrong way. You're looking at this hypothetical conscious/sentient washing-machine as a washing-machine that happens to be conscious/sentient; I'm looking it at as a sentient/conscious being that happens to be a washing machine.

The issue here with me is that it this hypothetical conscious/sentient washing machine would be conscious and sentient. Therefore, I would feel that destroying it would be unethical.

I would only say it's okay to kill/destroy a sentient being if it posed an imminent threat to you. For example it's okay for me to kill a person trying to kill me. However, otherwise, it's immoral.


INRM

 

[mimada]

INRM, I approve of your sense of ethics but I don't know what you are suggesting we do. If technology progresses as it has and at some point in the future an artificially sentient being becomes possible, how will the refusal of discourse effect change? I am not suggesting that anyone on this forum will create an artificial sentient or even provide hints to such a construct but the discussion of the subject will educate or at least broaden our perspectives. If you have some criticisms of rocketdodger's thesis, please share them with us.

As to the ethics of creating an artificial sentient, perhaps you can open another thread? I would certainly be interested in your insights.

 

[INRM]

Mimada,

INRM, I approve of your sense of ethics but I don't know what you are suggesting we do. If technology progresses as it has and at some point in the future an artificially sentient being becomes possible, how will the refusal of discourse effect change?

I suppose you make a valid point, though I personally believe that people should simply practice some self-restraint and avoid developing an artificially-sentient being. Not every idea needs to be entertained and embraced.

I am not suggesting that anyone on this forum will create an artificial sentient or even provide hints to such a construct but the discussion of the subject will educate or at least broaden our perspectives.

I hope you're right.

If you have some criticisms of rocketdodger's thesis, please share them with us.

I think the best way to explain a sentient capability as possessing the following characteristics

- Sensory Data Collection
- Sensory Data Integration
- Memory Storage
- Feedback loops between memory and sensory data, both integrated and non-integrated
- Inferential reasoning capability based on sensory-data, and based on memories already collected

I'm not sure about the last item being necessary for sentience itself, but an intelligent sentient being would have to have these characteristics.

As to the ethics of creating an artificial sentient, perhaps you can open another thread? I would certainly be interested in your insights.

I suppose it would be a good idea to create a dedicated thread, I probably should get around to that, but I should note that in replies 72 and 83 I did explain my ethical views to some degree


INRM

 

[rocketdodger]

First of all, I would like to say that I applaud your efforts. You seem to have done a considerable amount of research in the formulation of your ideas. I would like to address an issue, however. In my opinion, you need to be more descriptive of your terminology. Even if you dismiss some criticism as being disingenuous, it seems others are misinterpreting your usage of particular words. For example, you are using "symbol" specifically to describe information compressed in a NN but CapelDodger has apparently interpreted it in the more conventional sense as a "object, picture, written word, sound, or particular mark that represents something else by association, resemblance, or convention." (from wikipedia). I know that you have entreatied everyone to read through your post completely before responding but it is the nature of humans to default to common usage definitions (this is perhaps an example of Bayesian networks in operation).

Yes I agree. I only use the term "symbol" because Hofstadter did 30 years ago when speaking of a similar construct he imagined might be found in the brain.

I am formulating a better description of my model that should satisfy more learned individuals such as FedUpWithFaith, and hopefully by doing so there will be an implicit level of formality in the explanation that will be sufficient to avoid these confusions.

I do not want to see this discussion devolve into a philosophical debate about what conscious is so perhaps you could instead describe the properties inherent in conscious and how it pertains to your thesis. Also, it may be helpful to differentiate between human conscious and non-human conscious. I make mention of this because you made the distinction in your thesis and because a comparison may further clarify the concepts in discussion.

A basic description of the properties I am interested in will take a little time, but I can try to provide one.

Until then, I can say that I make a distinction between human and non-human consciousness because endless numbers of threads have devolved into uselessness over the stupid argument about why entity X is not conscious because it doesn't love, or write poetry, or any of the stuff only humans can do. Yet, when prodded, individuals admit that they think dolphins, or dogs, or apes, or even squirrels are probably conscious.

And furthermore, many of us (myself included) contend that the fundamental requisite of all types of consciousness is just some kind of self-referential information processing. Whether that is true or not isn't really up for debate, because that is just how we define it.

But nobody who knows what they are talking about would claim they have all the answers to human consciousness, which is actually what most laypeople mean when they speak of consciousness anyway -- else they wouldn't say things like "that toaster can't be conscious because it can't discuss the poetry of Robert Frost."

So at this point I am drawing a line and saying you know what, I don't really care if you accept that I consider some thermostats are conscious, because that doesn't mean anything anyway -- I don't think they experience the world like I do, and that is all that matters. And if you want to argue about something, don't be stupid and waste everyone's time arguing about whether a thermostat should be called X or Y (because thermostats are very well understood entities, so why would you argue about what they should be called? If I want to call it X and you want to call it Y, so be it, we both agree on exactly what happens in a thermostat, right?), be smart and lets argue about how stuff like human emotion arises from neural activity.

In other words, I want this thread to be about exactly how features of human consciousness come to be. I don't want to hear "it is a type of self referential information processing" because that is trivially obvious and doesn't give any details. I also don't want to hear "you will never know because subjective experience can't be explained by science" or any of the crud that flies around in the R&P forum. I want to know what kind of neural activity -- and what kind of abstract information flow that might correspond to -- is at the heart of our experience, responsible for our love, our hate, our thoughts, our dreams, our pain, our pleasure, and yes our own self awareness.

And the best way to do that is to come right out and say look, I don't give a darn about the philosophy, or even the general science -- this thread is about details of human consciousness. Does that make sense?

 

[rocketdodger]

I suppose you make a valid point, though I personally believe that people should simply practice some self-restraint and avoid developing an artificially-sentient being. Not every idea needs to be entertained and embraced.

I sympathize with your concerns, however misplaced they may be (since I am in no way capable of such a feat, and won't be any time soon), but if it is any consolation my primary interest in this subject has more to do with augmenting our own consciousnesses rather than creating and playing god with other consciousnesses.

 

[mimada]

...And the best way to do that is to come right out and say look, I don't give a darn about the philosophy, or even the general science -- this thread is about details of human consciousness. Does that make sense?

By the tone of your reply, I fear that you have misinterpreted the intent of the second paragraph of my post. I wasn't asking you to defend your reasons for making a distinction between human and non-human conscious. I was suggesting that a comparative study between, say human and animal conscious may provide some perspective and perhaps be useful in clarify aspects of your framework. I will forward some assertions to illustrate my point:

I believe that consciousness is dependent upon an organism's capacity to form abstractions. Humans, having the highest capacity for abstraction in the animal world, are sentient. While all BNNs are able to form an abstraction of their environment, organisms such as insects appear to function at a mechanical level and are not sentient. The quantity of neurons appears to be a factor for the capacity for abstraction and thus sentience. However, in larger organisms, the ratio of brain to body mass appears to be an important factor in the organism's capacity for abstraction.

The preceding paragraph has some highly contentious points that require further exposition. I will attempt to address these in a later post. Regardless, please understand that I support your project and only wish to help improve it as I find consciousness and AI to be fascinating topics. My apologies if my posting appeared rude or harsh.

 

[rocketdodger]

By the tone of your reply, I fear that you have misinterpreted the intent of the second paragraph of my post. I wasn't asking you to defend your reasons for making a distinction between human and non-human conscious. I was suggesting that a comparative study between, say human and animal conscious may provide some perspective and perhaps be useful in clarify aspects of your framework. I will forward some assertions to illustrate my point:

I believe that consciousness is dependent upon an organism's capacity to form abstractions. Humans, having the highest capacity for abstraction in the animal world, are sentient. While all BNNs are able to form an abstraction of their environment, organisms such as insects appear to function at a mechanical level and are not sentient. The quantity of neurons appears to be a factor for the capacity for abstraction and thus sentience. However, in larger organisms, the ratio of brain to body mass appears to be an important factor in the organism's capacity for abstraction.

The preceding paragraph has some highly contentious points that require further exposition. I will attempt to address these in a later post. Regardless, please understand that I support your project and only wish to help improve it as I find consciousness and AI to be fascinating topics. My apologies if my posting appeared rude or harsh.

You don't need to worry about apologizing, ever.

I don't agree with your terms -- since I consider an active insect to be conscious -- but I hope you can appreciate my sentiment that whether or not we agree about insects being conscious has no bearing on whether an insect is conscious like a human since they clearly are not.

And I do agree that a comparative study would be fruitful.

 

[Ichneumonwasp]

RD, I applaud you an a terrific attempt to put into words something that is very difficult to describe -- largely because we have limited vocabulary for it and the words we use (as you have experienced yourself) carry so many connotations that we are often stuck in equivocation land.

My own thinking is extraordinarily similar to yours, which is why I am going to point out the following issues -- some of the many issues that bother me.

1. Your initial description of consciousness as reasoning is fine, but I think it will cause no end of problems because unconscious neural processing is also based in the same sort of reasoning -- I think you might need something more explicit about what separates conscious and unconscious neural processing from the outset or you might want to qualify that initial statement a little further. Many folks have a tendency to jump on any sentence with which they might disagree on some level and argue that necessarily every part of your work is therefore useless.

2. While your discussion of 'lower level' sensory processing is very good, you might want to add something more about 'higher level' or downstream processing to the effect that the reasoning 'there' is not so much a 'decision' as a hypothesis. You basically get this point across in your discussion of seeing apples (I think you said something like 'when we see something that looks like an apple'), but it might help to be a little more explicit about this issue. We don't so much pattern match -- thing out there, pattern match to template of apple in head -- as hypothesize what we see downstream. That is why visual illusions are so easy to elicit.

This fits in very well with the idea of higher order visual percepts serving as symbols.

3. While I agree that emotion -- and please throw in motivation as well -- is a form of reasoning as you have defined reasoning I think you may need to be a bit more explicit about this area. Emotion is not the same sort of reasoning as say -- input of several dots along one plane of visual field = line detection. Rather, it seems to me that 'feeling' and emotion are difficult to define precisely because they are so fuzzy. A feeling seems to translate into some type of mild behavioral tendency that in some sense 'tags' an event or process. The experience of seeing 'red' seems like a color percept in association with connections both declarative (linguistic symbols of 'red' and its definition) and emotive. One of the very hard problems defining these words is any attempt to explicitly explain what we mean by a 'feeling'. Most folks seem to want to argue that feelings are not definable, not operational, unable to be represented by neurons or by artificial neural networks.

Calling feelings 'reasoning' is, I think, an important first step, but we must move further down the line. What sort of reasoning? By this point in 'experience' we are not talking about something analogous to simple line detection, but entire networks of neurons that provide some sort of 'push' for current or future behavior as far as I can tell. Perhaps you have another way of viewing it?

I think it is also vitally important when discussing this - along the lines mentioned earlier - that other animals possess the same or similar networks for feeling and emotion as us. In fact, that is one of the primary reasons why feeling and emotion have been relegated to the back seat of the bus -- well, that and the behavioral revolution from early last century -- so that we currently know so little about them. We focused so mightily on higher cognitive processing like mathematics and other form of syntax because that seemed to be what made us human as opposed to 'those animals'. Now our ignorance of these processes -- emotion and feeling -- has been hijacked by the 'you can't explain consciousness' crowd. That is one of the big explanatory gaps that we must traverse; so that is why I think it worth your while to expand on this issue further.

But overall a terrific start.

 

[mimada]

I don't agree with your terms -- since I consider an active insect to be conscious -- but I hope you can appreciate my sentiment that whether or not we agree about insects being conscious has no bearing on whether an insect is conscious like a human since they clearly are not.

I see your point. Since by definition, consciousness is "self-referential information processing" and insects, even with a relatively small number of neurons, still perform this function. I think that you should include this example of consciousness in the description of your model as it would dispel the interpretation (and hopefully deflect the argument) that consciousness requires the level of "self-awareness" that humans exhibit.

 

[rocketdodger]

More information! Due to a number of critiques, as well as some additional information being brought to light, I am going to better define, refine, and supplement my original posts.

The first point of contention was my definition of “reasoning.” I was not explicit in explaining that I am not speaking of what most individuals think of by reasoning, which is the higher order reasoning that humans use to solve logical problems (the kind of thing you might have to do as a homework problem in a philosophy course, or something).

So let me be clear – by “reasoning” I mean nothing more than a process of inference. Now, inference itself is the process of generating new facts based on existing facts, and again, I was not clear enough in what I mean by “fact.” I don’t mean a “fact” in human language, although such a thing is mathematically reducible to the definition I am using. Instead, I am using the notion of “fact” used in Artificial Intelligence, which is simply a piece of information that an entity can use to make decisions. I called this a “relation” between “percepts” or other “facts” because those are the only two types of knowledge that an entity can possess and all knowledge is merely a type of relation (in other words, percepts are also facts as well, but it is useful to distinguish them from other facts). Let me explain this as well, because some individuals have been confused.

If an entity is defined according to a system, then it is trivially true that all information processing that is part of the entity takes place within the system. Furthermore, we can clearly distinguish between state changes that take place outside the system (the rest of the universe) and those that take place within the system. If the information processing within the system is to reflect state changes outside the system (I.E. if the internal state of the system is to change as a result of the external state change), information needs to cross the boundary of the system.

When that information crosses the boundary of the system, the initial state change within some relevant portion of the system – that occurs because of the information coming from the outside world, so to speak – is what I (and the rest of the AI community) label a "percept."

Every other bit of knowledge that an entity might possess can only come from within the system. That means it was either built in to the entity when it was created (this type of knowledge is what most people call “axiomatic”), or the entity inferred it from previous knowledge. Note that the only advantage to making the distinction between percepts and inferred knowledge is that it simply makes the mathematics nicer by acting as a recursive base case I.E. recursions of a question like “where did you get this knowledge” will eventually end on a mix of axioms and percepts.

For example, suppose we have a robot with an inference engine (machine reasoning system) that includes the rule “If there is an obstacle in front of me, I cannot proceed forward” (it wouldn’t be encoded like that, but I wanted to make it easy to read). Further suppose the robot is rolling along and then there is an obstacle in front of it. When the forward sensors – laser depth finders, video cameras, sonar arrays, whatever – detect this obstacle, they will communicate with the rest of the robot software and somewhere within the program a state change will occur that represents the change from “there is no obstacle in front of me” to “there is an obstacle in front of me.” This is a percept – direct information about the outside world – and it is also a relation and thus a type of fact (as explained in the next paragraph). Of course this is simplified, but the definition still stands.

Once the robot has this percept in its knowledge base, the inference engine would reason (infer a new fact) that the robot should stop. There are many ways to implement such inference – such as the chaining I mentioned in the OP – but the end result is that the facts F1 = “I should stop if there is something in front of me” and F2 = “there is something in front of me” lead to the inference of F3 = “I should stop.” Note that all three facts are relations – the first is between two other possible facts (ROBOT, OBJECT == SHOULD STOP), the second is between the world and the robot (ROBOT == OBJECT IN FRONT), and the third is between the robot and a behavior of the robot (ROBOT == SHOULD STOP).

So I hope that clears a few things up. What I call a fact is merely a relation between bits of knowledge – which itself becomes another bit of knowledge -- and the process of generating such relations from existing ones is called inference, or reasoning.

Furthermore, note that while all inference is conceptually the same – going from existing knowledge to new knowledge based on some algorithm – the actual operation can be very different. So far in this post I have been discussing straightforward deterministic logical reasoning – if a certain set of facts exist, then a new fact will be generated, and that is that. But this isn’t really how neurons work, and I should have made it more clear that I was using the initial examples as just a conceptual framework.

Another type of inference that is closer to what neurons might do is Bayesian inference using a Bayesian network. This can be thought of as looking at uncertain knowledge and then deciding how probable a new bit of knowledge might be. For example, the robot would instead have knowledge such as F1 = “If there is more than X chance of an object in front of me, I should stop (with possibly Y probability, although that makes no sense to a human since they always want me to stop),” and the percept would take the form of F2 = “There is Y chance of something being in front of me,” and the inference would be F3 = “I should probably stop.”

But even that isn’t really what all neural networks seem to do, for reasons I will explain in another post. So I am going to introduce the term “Associative inference,” which in my mind should cover the gap. In associative inference, the generation of knew knowledge proceeds by associations between the relations between bits of knowledge rather than hard links, and is implemented by having associative networks be the input for other associative networks. I will get into this further when I discuss associative memory.

 

[mimada]

For example, suppose we have a robot with an inference engine (machine reasoning system) that includes the rule “If there is an obstacle in front of me, I cannot proceed forward” (it wouldn’t be encoded like that, but I wanted to make it easy to read). Further suppose the robot is rolling along and then there is an obstacle in front of it. When the forward sensors – laser depth finders, video cameras, sonar arrays, whatever – detect this obstacle, they will communicate with the rest of the robot software and somewhere within the program a state change will occur that represents the change from “there is no obstacle in front of me” to “there is an obstacle in front of me.” This is a percept – direct information about the outside world – and it is also a relation and thus a type of fact (as explained in the next paragraph). Of course this is simplified, but the definition still stands.

Once the robot has this percept in its knowledge base, the inference engine would reason (infer a new fact) that the robot should stop. There are many ways to implement such inference – such as the chaining I mentioned in the OP – but the end result is that the facts F1 = “I should stop if there is something in front of me” and F2 = “there is something in front of me” lead to the inference of F3 = “I should stop.” Note that all three facts are relations – the first is between two other possible facts (ROBOT, OBJECT == SHOULD STOP), the second is between the world and the robot (ROBOT == OBJECT IN FRONT), and the third is between the robot and a behavior of the robot (ROBOT == SHOULD STOP).

Thank you for that post.

In regards to this robot, would you consider it conscious? If not, what needs to be added to the construct to meet the criteria?

 

[mimada]

I think the best way to explain a sentient capability as possessing the following characteristics

- Sensory Data Collection
- Sensory Data Integration
- Memory Storage
- Feedback loops between memory and sensory data, both integrated and non-integrated
- Inferential reasoning capability based on sensory-data, and based on memories already collected

I'm not sure about the last item being necessary for sentience itself, but an intelligent sentient being would have to have these characteristics.

Your bullet points are relevant to the topic but could you elaborate? Rocketdodger does address sensory systems but you have made a distinction between collection and integration. Is there a reason for this? Please correct me if I'm wrong but, your last point appears to be rocketdodger's thesis. I am assuming that you that feel the remaining two points are essential for the model but are not adequately addressed by rocketdodger. I agree with you on the topic of feedback loops but how would you reconcile it with the model? I appreciate your contribution to this discussion and only request clarification.

 

[rocketdodger]

In regards to this robot, would you consider it conscious? If not, what needs to be added to the construct to meet the criteria?

Yes absolutely I think it is conscious.

Primarily because the robot could also be in a state where it was "on" but not functioning fully, and we could call it "unconscious" rather than "dead." So for the robot, consciousness is defined by it's behavior when fully functioning.

 

[rocketdodger]

I agree with you on the topic of feedback loops but how would you reconcile it with the model?

Associative memory, and associative reasoning. I will post about this later today, but the short explanation is that you provide a network with an initial state and it will settle into the closest local minima of an "energy" metric (called an "attractor" state) via feedback.

It is really fascinating and I can't wait to talk about it -- mainly because such networks are clearly a major component of our consciousness.

 

[mimada]

Yes absolutely I think it is conscious.

Primarily because the robot could also be in a state where it was "on" but not functioning fully, and we could call it "unconscious" rather than "dead." So for the robot, consciousness is defined by it's behavior when fully functioning.

Very good. This is consistent with your definition.

I would like to use this hypothetical robot as a reference point for further discussion.

If this robot had no sensory apparatus and no prior exposure to its environment, would it be conscious?

 

[rocketdodger]

Very good. This is consistent with your definition.

I would like to use this hypothetical robot as a reference point for further discussion.

If this robot had no sensory apparatus and no prior exposure to its environment, would it be conscious?

No I don't think so. Because it would have no notion of "self."

 

[mimada]

No I don't think so. Because it would have no notion of "self."

So can we establish then that conscious is systemic? That it cannot exist without a connection to its environment?

ETA: My intent in this line of query is to establish properties of consciousness. This is to enhance its definition beyond SRIP.

 

[rocketdodger]

So can we establish then that conscious is systemic? That it cannot exist without a connection to its environment?

ETA: My intent in this line of query is to establish properties of consciousness. This is to enhance its definition beyond SRIP.

Yes, although with the understanding that "environment" is simply some other system besides the conscious system that the conscious system can extract properties of itself from.

In other words, some solipsistic scenario works just fine as well -- the important feature is that the conscious entity is able to inform itself about changes in its own behavior.