Adobe Flash plugin required for showing the video here. Works in Linux, Mac and Windows. You can also download the video and play it locally in e.g. VLC player.
Adobe Flash plugin required for showing the slides here. Works in Linux, Mac and Windows. You can also download the pdf and open it locally in your preferred PDF viewer.
Prof. Heikki Hyötyniemi
AS-74.4192 Elementary Cybernetics
Lecture 10: Special Challenge: Cognitive Systems
Helsinki University of Technology, 3.4.2009
(v.2009.04.10, only a rough machine translation, not cleaned yet!)
[0:00 / 1]
Well, initiated ago.
This time we will be the subject of this most difficult, complex system.
At least, some say that kognitiojärjestelmä, or the human mind, the human brain, is the most complex of all systems in the universe.
This is the ultimate challenge, to complex systems research seeks to answer.
And on the other hand kognitiojärjestelmä is the mechanism through which all of what we understand, that is, all our complex systems theories and others, will have to go.
In that sense, it is quite reasonable, if you will be able to understand a little of this mallituskoneiston activity.
[1:00 / 2]
It is considered now that neokyberneettistä perspective to this.
And once again have to say that this is not in any way the only correct interpretation of these things.
Here it is assumed all things neokyberneettisten through the panes.
That is, however, held that history through.
One could say that behaviorismi was the first approach to cognitive sciences, that is considered the input-output behavior.
Eli was one of experimentation and it was considered how the behavior will change, for example.
Typical are these knee-jerk reaction when the dog trials, and others.
Then kognitivismin, and Constructivism through, came to the present time -- kognitivismissa focused more on these internal mental structures, and all kinds of constraints tämmöisiin what the mood is.
At the moment, you could say that this sort käyttäytymispohjainen approach is perhaps the most attention to the most receiving.
For example, robotics side, we talk about very much käyttäytymispohjaisesta robotics.
And, fun, maybe this is the fact that we could say that this sort käyttäytymispohjainen approach, then surely it is time to close mares behaviorismia way -- or in some way, this circle is closed.
Has it achieved anything in between -- no, it is why this behaviorismi or käyttäytymispohjaisuus of this sort is currently very popular, it is of course due to the fact that the methods and tools are developed.
Elikä computers are available, and then the other, neuroverkkoalgoritmit and others who have data oppivia methods.
Methodological advances have brought pressure on the here and the cognitive sciences, that cases may have to be examined in this new angle of vision.
So, in some way, because the development has taken place, then this circle is probably not quite a return to the former, but this is a mares spiral case.
And, this neokyberneettinen review, one could say that it is precisely an example on issues like spiraalimaisesta approach.
Instead of merely examining the environment impact, namely the behavior of the environment through both central role is also an internal structure -- that is someone in the middle of these two.
No other, here is an interesting feature, that actually the oldest most cases what the cognitive sciences in modern sense, it is said, so they have these Immanuel Kant ideas, and he pretty much thought of in this same way, 200 years ago.
After all, he said that -- or she, inter alia, said that environmental or habitat will not be able to think of these structures for what is around, but rather to set arbitrary limits on what the environment is detected.
So, I forced my structure for this observation.
This is a very profound observation, and remains correct today.
[5:23 / 3]
Now go neokyberneettiseen this aspect further.
This is again a loop of this sort it -- we started these reviews neurons scrutiny, and now returns to the back of these neurons.
Then neurons focused on examining just that, what kind of data analysis, what kind of compression, these neuronijoukot able to make the data, but it is now the challenge is that how could these neurons to explain the basis of, for example, symbols, concepts and categories emergoituminen, who are clearly at a higher level.
They are emergente, or emergent level of concepts, these concepts.
They are not direct observations can be -- or directly from the findings will not be able to explain.
The other, then tämmöisten symbol systems is built on top of this intelligence -- it is your turn emergenti level.
Mere symbols is not intelligent to be explained.
And even if it is älykäskin system, so the awareness is still a separate emergenti level.
Since the mere intelligence, which would not in any way guarantee that the system should be aware of.
Well, all of these some of this now thrown on issues like guesswork, here neokyberneettisessä frame.
Whatever the top I could say that because all of us, after all, is based on data, on issues like iterative review, so many of the traditional cognitive sciences or philosophy of mind problems can be avoided.
Because traditionally in the ruvettu although some concepts to define the other terms, lead to the so-called hermeneuttiseen circle in which the conceptual definition of the concepts will ultimately produce more than a vicious circle.
Now, after all, things at some stage revert back to a lower level inputteihin, that is, after all, some of the observation data sets, perhaps.
And the second thing relates to this particular awareness of the examination.
Often informs have to think that in order for consciousness to be explained, must have a machine, which interprets the lower level of the observation process.
Well, then have to be one of the upper-level homonculus again, which explains the lower level of activity the observer, and also here we get on issues like the end, the dramatic continuum, which does not ultimately explain nothing.
But again this kyberneettisessä frame may be, at least intuitively think that we have the potential to address issues like the dramatic landing of the problem.
[9:04 / 4]
Well, these phenomena kognitiotieteellisiä examined the other hand, the cognitive sciences -- where issues are examined analytically, starting with the existing cognitive and intelligent system, and attempt to understand its operation, ie the look, as if through the psychology of human mental activity.
Another traditional approach is the approach to artificial intelligence -- that it comes konstruoimaan schemes, which try to implemented intelligent behavior characteristics, and then deemed that it would be achieved in something interesting.
These are the two mares mutually supportive approach, which so far are not very well able to bring his face.
The basic problem is that then when you look at conceptual level, concepts and symbols at the level of things, so is not very easy to skip this gap here on issues like data-based analysis, on the other hand, if the source of the data-based review, as the artificial intelligence in some directions -- for example, neuraaliverkkotutkimuksissa when I start to build a data forward -- so it is very difficult to skip a level concepts.
Well, now may be able to understand that if the concepts in terms of data emergoituviksi, but also kyberneettinen approach can not say something about the concepts emergoitumisesta.
And the other thing is, of course, this, that when it is very korkeaulotteinen data, the so-frame problem, which is the traditional artificial intelligence problem -- it is that if there is any logic system, which makes reasoning of some of the known facts to justify, so when it is not known in the overall situation, so lacks common sense these systems are , always.
They can not understand this whole, on issues like the logic of systems.
But if kyberneettisessä system, all these emergent phenomena are based on a very large number -- after all, all the measurement data which is available -- at least in one sense, this is a real-world understanding, or the real-world data, at least offer the deduction as well, then.
Other issues like the traditional artificial intelligence problems such as aivoproteesin paradox, and then the Chinese room paradox, which he Searle said.
These may not be for you too much if you say you are not the artificial intelligence research have, but very quick to say that this aivoproteesiparadoksi comes from the fact that if we replace the brain one at a time, neurons in one of the artificial computing element, which fully implement all of the phenomena of this sort by the neuron to implement, should it not have it stage when all the neurons are replaced, so this is it, after all, do not agree with what the original brain did?
Well, this is very difficult to accept certain artificial intelligence researchers who do not agree with a strong idea of artificial intelligence at all -- who can not accept that the machine would never be able to do what a man does.
But in this kyberneettisessä frame can think of that is pretty much the same to what they are below the engine -- the most important thing is it only the dynamics, and issues like these attraktorit there in space, whether it be a data-space, or the idea for the space, so the key is just that balances model -- and the sense of this sort "arbitrary" in quotation marks can be really implemented even outside the brain.
[14:20 / 5]
Here are a few words of the traditional arguments for what side of artificial intelligence have been made, ie it is very conflicting views as to what the artificial intelligence can be achieved.
Pretty much of this debate ärhäkkyys due to the fact that this thinking is, after all, the man's own territory.
When the machines started to do any other human work, it is somehow thought that this mental activity is the man's own, which now has no reason to confuse the machines.
But even if we could take the artificial intelligence, so I think that is not too much to fear in the sense that we remain unique, yes.
No such return to later for the last time, these philosophical questions.
But really this sort more rational argument for artificial intelligence is this argument which claims that the human mind can not observe the human mental activity.
Or can not understand human mental activity.
Eli needed a more powerful machine that can understand that our minds are needed, and then the human spirit is not strong enough to be a machinery to do so.
The second aspect is of this sort then extreme optimism, with the thought that perhaps 20 years after the computer is so fast that it can -- even if it would be [sarjallinen, row] -- so is able to take the same amount of arithmetic operations per second, as the brain at the moment, or in the human brain can present to conduct.
So, automatically when the ability is at the same level, so it is the smart machine.
But, that will return to what we discussed earlier that even if an extremely fast operation of something, so it does not add to the intelligence, if you do not do the right things.
Artificial side, it's time to this peculiar phenomenon, that in almost every stage, it is promised that after 20 years things have to be solved -- this has already begun in the 50th century, and always has a 20-year goal is shifted 20 years forward, at some point.
[17:12 / 6]
Now, on this occasion of this sort will be intermediate between the two äärimmäisyyden between this sort directly insinöörimäinen words approach, and says that something interesting is probably available at a time when it comes to issues like the design of interesting structures to apply this cognitive problem -- in practice the measurement data are replaced by the computer data of observation.
It is formally just a small step, but conceptually it is still a large leap that, if the exchange on issues like sensors aisteiksi, or aistimet aisteiksi.
Pretty much on the assumption that, indeed, iteration and repetition of things, when it is made sufficiently high, and is on the E-operator, so we emergoituu something that direct data can not be seen.
[18:27 / 7]
But if I start this älykkyttä now in this way to seek, first there is perhaps reason to go through these ideas that what really should be aiming for.
Here is the interesting fact that the traditional intelligence or artificial intelligence, ie the Turing test -- it is based on the fact that the thought that the machine is intelligent when it is able to mimic human behavior in the way that an external observer can not distinguish the machine behavior in humans behavior, which in principle are expected to smart.
This is an extreme interpretation of behavioral and not be in any way interested in the fact that the manner in which the mind actually works.
In that sense, this starting point -- you can even say that it has resulted in artificial intelligence research in some way wrong raiteillekin.
Perhaps because of interest in the artificial intelligence, it is this sort universaaliäly -- or artificial intelligence in this sense, Turing can not reach before able to take the machine all of the sensations which a man may be, in this kehossansa.
But universaaliäly, which adapts to the sensations it is what each moment will be able to get, so it may have been feasible, however -- this sort of universal approach is easier to take than this sort of human specialist perspective.
Well, this definition of intelligence could also be argued, in a particular way, it is one of the emergent phenomena, which are almost by definition, are fleeing the definition of all businesses -- because at that stage where you define intelligence, so the stage can you also be implemented by a machine which carries out its definition.
For example, the field of artificial intelligence is very much on issues like applications, and a chess game is a very typical example of this, that first of all that was defined at the stage when the computer is able to play chess wisely, we can say that it is intelligent.
However, in the stages where it really was possible to carry out, ie the computer played very good chess, so it was found that the targets move.
That -- because we understood how the machine works, so it no longer corresponds to our concept of intelligence.
Since then, in particular, that no, yeah, but until it wins a good player, or until it wins the world champion, so it does not really have the intelligence, and at the moment when the world champion has been won, the goal is again moved forward.
So, in this way that can not be traced to this intelligence -- it has to be approached in any other way.
And almost the same manner as these systems, when identified, the system is an entity which can outline the system, but also the intelligence are such that ihmisäly or intuition to perceive intelligence.
You could say that this is one definition, and if you are approaching the point of view on this matter, then perhaps the most logical solution is to ask the experts, the most älykkäimmiltä people how they define intelligence -- what their intuition is the matter?
Asked some Mensa members on this matter.
Well, they älykkyyskokeissaan measured intelligence.
And, after all these intelligence tests, a measure of how well you can detect havaintodatassa characters.
So, do you have a few song a few examples -- you should test for intelligence to be able to detect that which is a continuum, or what is common to them all for the pattern, and it then concluded the string pattern will continue.
This definition is only the sad side, this sort of artificial intelligence is very easy to implement -- I have done on issues like the program that can perform intelligence,
[23:34 / 8]
or take an intelligent behavior in this regard.
In other words, if we have the world so narrow that it can be reduced to issues like, for example on issues like the grid, then all these kinds of associations and Avatars are computer-controllable, even better than the man under control.
For example, you can watch this here in the top row, that what semmoinen feature is common to these three primary kuviolle, what does this fourth graph is, elikä what is the simplest definition of the character that separates the three-quarters of this?
Well, semmoinen solution that machine for this found, is that this is just one of this sort alkeishahmo, in this picture, if people in the other two tracks.
Here then in the next row is the same question, namely, the first three patterns among the fall, and the fourth does not belong to them.
Well all three of these the first is a horizontal line, and the fourth is not.
This particular engine found -- micro-second.
[25:03 / 9]
Well, maybe we should take a little wider perspective and see how that in some other cultures than western intelligence has been defined.
Extremely different point of view it is really here in the Eastern religions, to a certain Zen-Buddhism, starting with the very opposite point of view.
So, intelligence may be, or the heart of human intelligence of the image to open operation, it may be just this sort of non-proseduraalinen thinking -- totally assosiatiivinen.
Only then when you can, all the ideas you think the expulsion, it is only when you have reached that stage of enlightenment in which a certain way you understand all the issues, but that you have to separately think about those things.
This is now, at least in my interpretation of this.
In other words, these, on issues like, zeniläiset koanit, are just examples of how people try to teach to the manner in which the human mind of thought through to produce paradoxes.
In other words, these Zen-master is the purpose of it, that leads students to the state which can no longer take any questions.
And the last here, particularly suited to this kyberneettiseen thinking that before the mares Zen absorption, so people can see the mountains and the people on issues like separate cases -- the stage when you begin to understand things deeply, then they somehow mixed up these things with each other, that does not completely separate, really no matter the world, everything is in some way a common -- but then when you have reached this zeniläisen target, so the stage again, these mountains and the people are separate, but one of the upper level, not on the everyday level of thought.
But exactly the same way as those in other koaneissa, so this course is not intended to start to think that what this higher-level thinking thinking now is, since then -- if you think of Buddha, then you need to kill the Buddha then you think, because it is the only the way forward.
[28:06 / 10]
Let's go now actually, just insinöörimäisesti things further, in other words that we said just now that in fact these traditional approaches, they are very confusing and will lead only to paradox.
As the starting point for the very neurons of this activity, and see what can be achieved.
And seek to interpret it then the end result of these theories kognitivististen framework.
The idea now is that there was no other than the number of neurons that are looking for information -- or apply for variation, applying the resources -- all are competing with each of the variation, and it is the neuron which is the least of all received this herätettä, it is the most active, it wants to most connect to any of the information source.
Now information is for example a variation in terms of these aistimista, ie it can be argued that at the lowest level, the lowest-level associations are those linking neuraalisia, or issues like sensory findings together.
Well, when this sort is the lowest level neuronaalinen level, some kind of network -- assosiatiivinen medium, as used in this mares name -- so that you can start to build more complex structures.
In other words, if there assosiatiivisessa online now are some elements activated simultaneously, so one of the next neuron level can start mallittamaan activity of this combination.
If this sort of activity is a combination of relevant, that it occurs frequently, or in the way that it is stationary sense meaningful or useful, relevant, then this neuron, it is a happy and more or less stabilizes to present a variation of this combination.
Then, when it eats this variation from the system out, so others will no longer see it as a resource, and are forced to compete for some of the remaining variation.
What is nice is the fact that this neuron that has now found its own role to play, so it becomes activation, which means it receives on the lower activation coming in itself, and this activation may be the indefinite hunger in the neurons start to turn to.
In this way, it can be argued that this sort, at least the lowest level of the Associations, the formation does not require any high-level control, in principle.
But exactly this sort tabula rasa-metaphor without an external guidance is sufficient to build the foundation's findings mares someone kind kyberneettistä whole.
Well it is now possible to think that if we, for example, hear a story, so how can it in this machinery then learn, or remember.
[32:08 / 11]
Well, here is an example of the fact that, for example, has heard the story of this sort, the mischievous boy throws a stone window and the window breaks.
Now, if you think that it is sufficiently complex kognitiokoneisto already existed, and lower-level assosiatiivinen media is available, then the information that refers to a concept of the window, this will activate the network of a particular set of neurons, which will become active.
And, now, or is the concept of the window to eat this activity, which defines the concept of the window.
But now this story, in this example, the evil son of a concept or a particular nasty son of a concept, which has been talked about for the past.
Now there are these, that a certain boy, referring to associations now active, on the other hand, the term malice, and so on, and these come together in this single neurons, or neuronijoukon, kidnap why, in a particular way, ie they are linked to rapid dendriiteillään, now that a number of which happen to be one time active.
Similarly, when the next will this mean that the stone is thrown, the boy has thrown the rock, this rock-throwing, it is this concept which is activated, that is just this certain rock-throwing is activated, and it is linked to the current package of the Associations.
Still higher up are then this mean that now that it is mischievous boy has thrown the rock, then this story is progressing here neuraalisella level in the way that this nasty boy throw the stone has its own, in some sense already in the symbolic, conceptual entity.
The whole story, still, when it turns to the evil of this boy throw a rock and a broken window, so it represents the deepest level of the story, or a single sentence, the whole.
Well, these are now well on issues like the associations of individual instances of expressions here.
If this never again appear in the associations of this package, so they do not become active again, and their activity level, its single aktiivisuuspurskauksen after descends towards zero, and this is the fact that they have to seek a new activity, ie, they have to seek a new concept or new things assosioitavakseen.
They tend to forget this story then.
If not, this story repeated often enough.
Or, if this is not in any way relevant, in which case they often enough, these are activated when these connections are strengthened.
In other words, if they are long-term relevance of these connections, then the neurons, this activity remains at a satisfactory level and it is remembered this story.
Conversely, if these are the general concepts, which -- for example, as if this window here, or here in the stone -- they can occur in very different contexts, jolloinka the coupling of individual associations, it is not semmoista further-reaching importance, and indeed in a higher level of a relevant category is someone's son or rock, or throwing, then these issues like the concepts -- just the stone, the boy, throwing -- they are activated in itself enough, often enough, but the above might not be formed any permanent structure now -- but when these assosioituvat well, or are activated in different environments, these stone, and the son of the concepts, so they eventually vajoavat relevant noodeina, concepts relevant to this network, in which they are probably available, but their connection to a childhood in a deeply felt experience of the stone, no longer able to return, but it is in some ways this semmoista the so-called kvaliaa involved.
So, it is very much on individual connections to old bottom experiences that are not eksplikoitavissa, anymore.
In that sense, all people are different in their concepts of composition in itself, but if they are the people here in the world objectively important concepts for them, to us, consists of interactive objective world, or at least inter subjective worldview.
So returning to that last time, yet, these tämmöisiin episteemisiin questions.
Well, this review will lead to the fact that we can immediately respond to some of the traditional problems of cognitive sciences.
And really just to this cognitive sciences and artificial intelligence to traditional gap between the problem, why is it so difficult to get from here the symbols for this level of data or the other way around.
[38:50 / 12]
Now is the harpattavissa over the novice and the expert knowledge of difference, or the presentation of the distinction, the problem of excess.
Traditional cognitive sciences the problem is that when a beginner learns -- for example, learn to play chess in the way that knows that the horse is moving this way and that, jerk moves this way and that.
He is not yet very far-reaching strategy to be able to build on these -- it is the experimental aikalailla this rule-based action, this sort of thinking novice.
Instead, the experts of this sort is a character type of thinking that sees a situation in mind, to restore it in some way, is able to reconstruct it in some of the former arrangements, and, for example on issues like openings and the other, they are semmoisia -- that is, a chess game to open the situations are certain types of categories of what the expert sees and is able in a particular way reconstruct this board, tämmöisten meaningful categories in mind.
This expert thinking is precisely the mares assosiatiivista, figures based thinking.
Traditionally, it is very difficult to combine these two -- or to explain how this sort of thinking beginner becomes an expert thinking.
How deklaratiivinen this sort of knowledge is changing assosiatiiviseksi knowledge?
Well, the recent story of learning, or a rock concept learning there is perhaps more tools for this learning, or understanding.
Traditionally, of course, when talking about expertise, so any talk of a high level of expertise the complex problem of perceptions, and perception, but one of the rock-like understanding of the concept, or a chair for the concept of understanding, it is of the human tacit knowledge, which is not usually able to adequately appreciate, until it start to take some robots.
For example, the concept of the chair, it is anything more than four feet and back -- a very long assosioituvat all the functions on issues like root, which is sitting a function is an integral part of the concept of the chair, and so forth.
[41:55 / 13]
Well, now when it comes to think that the concepts are based on real kyberneettisiin or neokyberneettisiin basic structures, then we can start to approach this issue from another angle.
So, what kind of structure kantilaisessa sense, this kyberneettinen machinery make it havaintoaineistolle?
If you have a few observations only exist, or a few data points, which presents, for instance -- one of the child, it is seen, for example, a few different dog, and a few of these dog concept, or visual observation, the preliminary sense, defines the child to the concept of a dog, and, indeed, even if it does not have seen so many dogs, so it is the dog's concept already exists, which is more precise then later when you get more dog findings.
But at the stage when there are only a few dog detection, so you could say that this self-concept structure is already in place, because this kyberneettinen machinery is forced to the data, that is, where is the data aligned to this concept of a dog on the number of neurons, so this interpretation is the same as if it were a large number of data, which would have led to the concept of a dog.
Eli is in some way from the top of view -- when it comes to interpretation, so neuraalirakennetta -- it can be argued that it is a virtual data, which is tuned in this material, or this, on issues like the distribution.
Since a certain way kyberneettiset these structures, they represent the data distribution.
As this data, the distribution of the image are a kind of virtual data, which is then distributed muokkautuu depending on when you get more material.
That what is the virtual breakdown of the data now, which all neokyberneettisille structures set?
It is based on just that, to these major and other, more practically, it can be said that the entry for the virtual breakdown of the data is Gaussista data, since it is precisely the Main, they represent Gaussisen data distribution for the best possible way.
The only thing we can for this breakdown of the findings, therefore -- or what is the best way to say, it is returned tämmöisiin Gaussian distribution, then.
So, what does this mean?
How do I now have to interpret these ellipsoidit -- that while this data distribution, mares Gaussista the distribution is interpreted korkeaulotteisessa in space, so its interpretation is of this sort ellipsoidi different projections.
What are these ellipsoidit now?
No such major ellipsoidin are now such a key then.
So, Main, these, they are clearly relevant to this interpretation, since it is that one of this sort, something to the concept of describing the distribution is a particular direction, a longer period of observation in space, then certainly it is one of relevanttiutensa.
[46:02 / 14]
This now looks just that, how can we interpret these pääkomponetteja.
If it is found that in a certain direction is a lot of variation of a concept around -- it is a relevant concept, but the concept has a lot of heittelyä -- both in the direction of this distribution is then elongated.
For example, if you have any concept of a dog, and then we have seen in very different sizes of dogs, however, so this whole concept in mind -- that the whole concept is a low-level concept, which is attribute to the concept of a dog -- the sense we have of this sort, this elongated, and then , for example, no one ruskeus-axis is then a second shaft, which direction is more or less of this variation.
Well, the dog's case, now this ruskeus is the time -- it is time for a large variation because of the white dogs and yellow dogs, and so on, in other words, this now is a very good example, perhaps not, but each of the random observation dog -- for example, here is this sort Fifi-dog -- so it now can say that if it has been said that this poodle Fifi now part of the dogs, so the child form, or to edit this dog then the concept of it.
It might go in here, this ruskeus-axis direction now, a little laveammaksi, as it is now this sort white dog now -- and whiteness will help to determine the dog then, too.
In other words, the Fifi is an exceptionally small and exceptionally epäruskea.
This center here in a certain way now, perhaps some kind of alisymbolista concept.
One could perhaps speak of miellerakenteista, or muistiedustuksista, or chunkeista, English speaking.
The upper level can we talk about categories, and the characters.
And then again, these, who virittävät this center in different directions, are issues like attributes, or else someone nyansseja types, or features.
This is time to close the conceptual spaces approach, which has been examined in the cognitive sciences -- for example, Peter Gärdenfors, Swedish artificial intelligence researcher, has presented on issues like the approach.
Semmoisenaan But this is not to be a very plausible model of our mind, the mental or the infrastructures -- of this sort a linear model.
[49:30 / 15]
Well wait and see before we go deeper into the matter, so looks a little interface to traditional kognitiotutkimukseen.
One could say that the prototype theory is very close to this approach, ie it is connecting, when replacing a number of these concepts in others, so it can be seen that they have points in these traditional areas of research.
Eleanor Rösch has a name that is pretty much the prototype of this theory, or in some way identified in the original proposal.
In other words, every concept is a prototype of this sort, such as the prototype of a dog, and then when you have different prototypes, it is then a model of mind in it.
No fundamental problem here is that, what is this internal structural issues like prototype -- if it does not have any machinery or the handling of the machine, so this is pretty much on issues like the wheels of conceptually, and, for example, in the artificial side when applied to this approach, it is thought that we can on issues like prototypes start to apply the reasoning, at the technical level, it was reached on issues like the Case-Based Reasoning model that just think that the various examples of past behavior, they are on issues like prototypes, and then claimed similarities in the case of new and old prototype inference, between a, which can then be leads to the fact that what should be inferred from the new situation.
Well, where does the prototype interfaces have been carefully defined metriikkaa, so the problem is the fact that in what way can these prototypes to determine the proximity, that is precisely the Case-Based Reasoning -- it is just problems with that if you do not have the internal structure of these prototypes, so not really possible to do very much.
Well, neokyberneettisessä frame, we really have metriikkaa, which is the basis of data determined, rather.
No philosophy, just half ago, is this sort Embedded Cognition embodied approach, which pretty much starting with the same thinking as that Rodney Brooks käyttäytymispohjaisessa this approach -- in other words, these cognitive concepts are meaningful only in the sense when they are connected the world, or ruumiillisuuteen, or brain.
So, this is however, quite modern, it is not the way out the end of kaluttua this matter,
[53:06 / 16]
in any field of research, not cognitive sciences, philosophy is not, nor neurotutkimuksessa and robotics, ie the field is quite a bit more open to new proposals.
And then the artificial side is a lot of issues like that other approaches, which have close access point to these recent reviews, including -- for example, the fuzzy subset is one approach to this sort of knowledge, particularly for modeling.
And then, semantic networks, which explicitly talks about the semantic neteistä, and semantic webeistä -- suomessa now on this e-word, now it really describes both of these things, but the semantic connection nettien specifically address such -- no, there will be a moment,
[54:06 / 17]
or this is just an example of this sort semantic network in this sense, the Web.
So, these are built for a long time before we had networks of semantic web in mind.
So, for example, defines this sort Fifille concept map in the way that it is an example of the dog, which is an example of the domestic animal.
And then Fifillä a feature -- no in this case it would probably feature a white, if it is of this sort in the dog.
But what of this dog in the core itself is, so this does not really get very far, because this is just the sort of traditional artificial intelligence, which -- in order to get a meaningful whole, must be reasonably small number of these real-world circuits.
So, Fifi, for example, defines koiramaisuuden through, and the White, and if it comes here to do reasoning, it is precisely this understanding of the real world, all koiramaisuuden features are absent, ie it is this framework the problem follows from this, issues like the review.
Now, in this framework kyberneettisessä there a data-space, is this really Fifikin in a certain well-korkeaulotteisen space section.
Some of these attributes to it Fifille is determined on the basis that we know that it is a dog, but some attributes that are explicitly seen it, so are determined to visual observations, such as whiteness.
Well, the fuzzy subset, this is perhaps a picture of it.
So, it can be argued that tämmöisten pet animal population is the most extensive set of, then the number of dogs has been since its subset, and Fifin -- or Fifi is then this series of dog embryos.
Well there is a little problem with these concepts sumeissa it, on the one hand, this ruskeus or whiteness defines also the Fifi, a dog.
And the fuzzy subset has a strong sense of theoretical construct that may be, for example, the brown or white in mass, it may include also this Fifin, although this Fifi only contribute a part of this valkoisuuteen, and contribute to the koiramaisuuteen, while koiramaisuus much smaller extent in this part of the valkoisuuteen.
So, this blur these osajoukkokäsitteissä means that when a series of embryos in traditional sense, is indeed an item, then this fuzzy subset in mind also the embryo is a collection which has a small part, to determine the ylijoukkoa.
This neokyberneettinen framework is the sense of this tämmöisten fuzzy subset concept implementaatio that all the concepts defined by each other -- they are also upper-level concepts, while later, then consisting of the coupling, you will notice that if Fifi is often active when the reference to dogs, but also koiramaisuuteen coupling comes from the upper the level of concentration, ie Fifistä, and then this one example, for its part, determining the content of the dog, my small contribution.
In other words, if koiramaisuus defines Fifin, but also defines Fifimäisyys dog, in a sense.
These are the traditional sense, logic, a set of traditional doctrine in mind a little obscure, these things, these fuzzy things, but the fuzzy subset of the framework may be just the subset of the traditional weaknesses of a variety of ways to avoid.
And when this kuuluvuuskin, it is not semmoinen one-to-one or 0-1 type of addiction, but it is a relative of this sort -- that is one part of the penguin birds, yes, but only to a certain sound, a bird, including the prototype is less than one twig on, for example, because of the flight performance has Penguin is so characteristic of those of other birds typically have.
[59:25 / 18]
Well, then if you think all these fuzzy subset on issues like terms -- for example, ruskeus has its own käsitteensä which define the brown objects, and koiramaisuus has its own käsitteensä which determine its characteristics, for example, ruskeus their part.
So, if the brown things, we are often the dogs in question, then -- you have to remember that these are all issues like these additive neokyberneettisessä this framework, these linear features -- in other words, this koiramaisuuden direction korkeaulotteisessa data-space, now can summautua this ruskeuden prototype, jolloinka dog-like brown area in this area has one of the maybe, while the brown dog is here closer to the dog prototype.
In other words, a certain way it is the most powerful at the time the current attribute is the self-concept prototype.
In other words, is not separable, or is not profitable to distinguish between individual category and the concept of the concept of attributes, because kyberneettisessä frame may be the internal action to obtain just the same frame, and then just this sort thing that koiramaisuus is both a category that feature, so it can respond much better than the traditional concept of the hierarchy.
If, for example, in a C-language to describe the dog is a pet animal subset, or Fifi is a dog instance, or one example of the dogs, so it does not in any way help you in that dog's definition of the concept, while kyberneettisessä frame, if we feel this Fifille a lot of properties, they contribute a small part, determine the dog's characteristics.
If the dog does not know any other examples, as these properties Fifin then inherited yes.
[1:02:06 / 19]
So, the concepts are given in the data-space, on issues like relevance attraktoreita.
If you have a data-space, often enough led to a certain point --
In other words, these characteristics, or attributes defined within the concept of orientation in the data-space.
[1:02:34 / 20]
Well, if you will now be defined in such a conceptual space or the semantic space, then noticed very quickly that there is simply no adequate description of the power on issues like design, for example, because if we think of dogs and horses, so they have certain features in common, for example, they are all animals that breathe, and they are up to both domestic animals, but at a very early stage, the child learns to distinguish between a dog and a horse from one another in such a way that it has a very precise limits of these two concepts between.
That even if a camel can be mixed with a horse, because they are a certain way, overlapping, so the dog at a very early stage, separated by horse.
In other words, these concepts are very much mutually exclusive, these two concepts, they are not linearly additive -- or even a horse and dog are both domestic animals, so it does not mean that they were very close to each other, psychological world.
Lineaarisuusajattelu of this sort does not work very well if you think the correct functioning of cognition.
We need something epälineaarisuutta, and proves that this sort multimodal data -- ie if there korkeaulotteisessa data-space, there are several issues like the concepts of clusters, ie the sensing material forces -- so they can be well described by this linear approach to expansion, that is just as harvakoodauksella what we demonstrated on issues like the selfish, or neurons neuronijoukon do.
In other words, some features are relevant only when joining a particular concept, and some others are joining the second term, but nevertheless, for example, this pet concept is relevant to you and the dog that the horse, and in that sense this distinct categories may also be common features -- and in that sense, this is harvakoodaus a very powerful approach that we can learn from these pet features when we learn something about the dogs, when we learn something about horses, and also when we learn something about the chicken -- but that this kanamaisuus now affect any horse.
In other words, they are the concepts of dog, horse, chicken, can share this one pet feature here harvakoodauksen frame.
[1:05:55 / 21]
This is still a multiple that what this harvakoodaus now in practice, then it is.
And, this harvakoodautuminen is an example of a way, that on issues like how the contradiction -- or reverse intuition come against this when it comes to making mares real world mallitusta.
Traditionally thought of as and when the neurons are linked to other neurons, so it is with the intelligence to grow, or has the opportunity to more -- or eksperti improves, and so on, but if you just be more connections, so it would mean that everything will change eventually as one on issues like möykyksi, on issues like assosiatiiviseksi möykyksi -- key The point is that an expert can distinguish between things quickly from each other, is precisely the harvakoodautuminen -- that is just part of feature is activated.
So, instead, that it would be the case that there is a lot of connections in different directions, so more at some point these connections gonna katkeilemaan -- that is gonna come to this harvakoodausta, that is only part of issues like features is really relevant -- although often in a certain complex of expertise required in connection with the some disturbing things here, so the expert can, however, to distinguish them from each other, are able to filter out irrelevant in some cases then, after all, completely off, sees clearly only those most essential features.
[1:07:56 / 22]
Here then is this that how can something such as muscle activity may then modeled.
[1:08:04 / 23]
We dealt with this already, this is Olli Haavisto thesis work, there a few times ago, and we said that we do not need any dynamic model of the head, or brain, kognitiokoneistoon still be -- but this state of mind that issues like the dynamic system, such as in this case the walker, is needed to right situation, know how to use the right controls, so it does not even have to take the dynamic, the brain, the dynamic adjuster, but can be confident of the environment itself, which is dynamic, it will deposit its system state.
If we can measure the muscle of all stations in each moment, then the role of muscles in itself tells what steps are being stage, ie it can be done purely static assosiatiivinen linking to the muscles, and then when the muscles work, so the real cause of the muscles that move to a new drive, and gravity leads the fact that this configuration is then another, and move to a new drive which in turn can be found here, here view on issues like the assosiatiivisena description.
[1:09:40 / 24]
Then there is an apostrophe in there Kohonen maps direction that -- Kohonen kartathan is a very different approach to mental mallittamiseen, ie it must be observed that the cortical there are various issues like operations, or different things are like kartoittuneet different points, ie close to one another on the agenda, are close to each other also the brain.
And if we think that these neurons are mares ylikuulumista, so we think that the Q-matrix, which was originally diagonal matrix -- meant that one neuron to affect only the self -- it can be argued that it diagonaalimatriisi is a little bit wider, which means that this Q matrix reflects a kind of naapuruussuhdetta between neurons, ie one neuron affects ylikuulumisen is also one of the second neurons, it follows from the fact that close to each other, often on issues like equal time to activate features will also be close to each other.
Well, it is difficult to prove, but if you go through this SOM's theory of the so-will find that it is todisteltu various experiments, that is true, this map idea seems complex cases then emergoituvan.
But compared to that Kohonen map really so here kyberneettisessä frame has an internal structure of these noodeilla, which is a bit of added value, yes, if I want to do something more than just a visual examination of this map.
[1:11:33 / 25]
Well, then could be examined in that the type of coupling points is tämmöisiin kognitivistisiin theories.
If you have kognitiotiedettä read so you know, that just this kognitivistiseen phase involves issues like mental activity and the constraints of the other, for example by the observation that man is this, a certain way, the channel width constraints.
So, one time may have in mind only a limited number of things.
It is famous for semmoinen [Miller's] paper, which states that the 7 + / -- 2 is the human mind as an active time of the maximum.
So, for example, are based on memory techniques, so if you for some mindless string of numbers to remember, so you can remember it as a way of dividing it into smaller groups, which groups a total of maybe then this seven -- if you these small groups you can find some sort of internal structure, which in turn is composed of on issues like 7 + / -- 2 to alirakenteesta -- for example, a long-telephone number to remember if it consists of some years, although the figures or some other date, because then we can restore it to some other memory elements at a lower level.
This kognitivismi enough to work, but it can be shared, this is limited thinking, this way at lower levels.
And now, if we compare this kyberneettiseen frame, as well, this works in the sense that if you are limited by the fact that the number of mares harvakoodattua features as time is running, so just the same way a certain number of connections at a lower level of muistiedustuksiin, coexist.
So, this harvakoodautuminen, and this kognitivistinen limited thinking is very much similar to, although this sort on the same footing -- would require a great deal semmoista problematisointia and reflection, and -- now says that only an engineer can do this a ridiculously equation very different between the issues, without taking a position on the fact that what matters is discussed earlier.
But in this neokyberneettisessä frame can focus on what is really relevant -- they are data, the facts I determine relevanttiutta their own way, [and] issues of materiality, it is nevertheless more important than some of the names of the theories.
[1:14:57 / 26]
Well, this issue brings relevanttiutta fact that I have worked with Pertti Saari Luoman, who is now a professor of cognitive sciences at the University of Jyväskylä.
He has been on issues like a chess-like ekspertiisin global scale an important scientist.
And chess is an interesting problem, in the sense that it is cognitive sciences to many problems well crystallized form detectable.
It is this ekspertiisin novice and the difference is here in a very compact semmoinen problem, which can then be investigated experimentally by far.
Here is an example of the fact that there have been 5000 shakkikonfiguraatiota, and they have been taught then issues like harvakoodautuneeseen map.
This is a very simple manner, coded chess, this situation -- it has been a 768-dimensioinen space -- a simple, in the sense that it is a chess board described in such a way that each box 64, the end is a 12-bit vector, where each bit, the vector means that a a particular piece is in the box, and when these bittivektorit now assembled together, so it will be this 768-dimensioinen data.
So, all shakkikonfiguraatiot may be uniquely present in this framework.
A lot of illegal situations also, but the hope is that this machine is able to find the relevant configurations from korkeaulotteisesta space.
This is now pääkomponenttimalli it.
Here is the beginning of the situation, and as and when the pieces reduces the board, as might be driven towards this zero space -- that is, none of the bit is no longer up to the description of the situation.
[1:17:08 / 27]
It turns out that the various openings, for example -- various shakille typical dress, for example, and the comfort of others, they are on issues like chunkeja, which typically occur.
Now, do not need to learn all the locations of buttons separately, if you now have 32 pieces -- no, four times eight pieces in all -- so this 7 + / -- 2 restricted, beginners means that is able to remember only a few checker position, but the expert, as these experts require chunkit has built up -- just in the form that certain issues like the military deadlock, and linnoittautumiset, and certain strategic patterns, openings, have their own categories -- so he can remember very well the whole chess board, after all.
[1:18:24 / 28]
Here is an example of one example -- that these have been studied in just the experts, proven mares situation five seconds, asked to reconstruct board, and now we taught them chess relevant situations to the machine, and showed the same board -- that was given it -- the thought of it in five seconds, it will be able to connect to this one-level description of the board, ie to find the links to the lower level shakkiorientoituneisiin chunkeihin.
If you have a copy of only those muistirepresentaatioita, ie short-term memory is one unit long, so may be remembered for so many pieces -- we find that there is a lot of detail incorrect.
As and when these become hienovirityschunkeja add, that are linearly additive, so will always be able to just more accurate.
For example, typically, this horse has the movement semmoinen that one begins to become more or less single neurons or single issues like a low-level description to propose, because it can vary, typically between two points.
But what is most interesting, is this the fact that these errors, what does this description does, they are very asiantuntijamaisia.
So, this was the one who attracted Pertti Saari generated particularly high, because this is seven or this is, frankly, three, or five chunkilla of the situation.
So, there blood when you are in this situation the right to see that this tower is in the wrong place.
Originally, it was that next to the king, now it is here in this.
This is a very semmoinen expert standard error, as this tower could be either here or here.
It is not very much affected by the fact that whether this is a relevant situation in a chess or not.
In this course, have to remember is that at the stage when it is reconstituted the board, so we have accepted only those reconstruction, as a result of the characteristics of the aggregation of and by doing it this sort threshold, ie a certain threshold level or threshold, above which the reconstruction only is accepted.
So, there are issues like a low-level access to it that this model could be based in one box, so it is very much of course, but in this particular situation, most have agreed to semmoinen reconstruction, which is kind of binarisoimalla then looks like this when it is projected here 768 -- dimensioisesta from space back to this shakkilaudalle ago.
[1:21:52 / 29]
And this is really interesting in the sense that it is able to mimic the expert versus expert behavior.
You could even mention the root of this yet, that one thing which created the island were interested, it is this that when you have done research in the right shakkieksperteillä -- so that the chess eksperti will be able to remember the chess these situations is not because they had a better memory, because if these pieces is put random order, the board, so essentially a chess expert is not able to remember those buttons positions any better than the starter either.
Only in the event that the situations are shakille meaningful, and even the good games -- that this is just important that chess eksperti learn chess by going through good games, no bad games.
One random player might suddenly surprise eksperti, when it is dropped from the opening out of the library.
[1:23:04 / 30]
Well here is on issues like more examples --
[1:23:08 / 31]
for example to do just that, when discussed last time, it Pyhäsalmen data, so this is an example of the fact that the experts are actually seeing in this way a rigid, wet and dry foam, and this is not semmoinen kategorisointi, which they would be used to make explicit, but they see the bubble surface someone on issues like, more or less subsymbolisen rating.
[1:23:41 / 32]
But we now that we wanted these experts determine the bubble sizes and läpinäkyvyyksien and all other surface of the bubble velocity and other characteristics, these soft sensors produced by the characteristics of the bubble in which the surface is now what class.
It turns out that this is our hand-coded piirteistin generated a specific period of time the result of mares.
There, in the range of the long term become such a manner that is more rigid and less rigid foam in it.
And when we did just purely datapohjaisesti of the same data in learning, so we noticed that these unidentified harvakoodatut features, they began to represent the fact of these -- at least in some sense, or to some extent -- these experts defined features.
This is the second of this sort a good pump bench expertise, because a certain way even though they are on issues like eksplikoitumattomia in some sense, these asiantuntijuus stuff -- that is one of the operator will have the hard way to learn the process in action, now -- it's a very relevant issue, therefore, now, and you think that this is an interesting problem there -- this is now a well-research yet.
And what this is interesting is the fact that if this data were made directly pääkomponenttianalyysi, so anything like this does not really been seen -- or perhaps the strongest feature, but everything else was just pure random data -- even if it was optimal, mathematical sense, optimally koodautunutta, so it did not have these direct connections to these expert categories, however, no longer visible in the pääkomponenttikoodatussa material.
[1:25:54 / 33]
Well, then this has been tried in some other applications --
[1:26:00 / 34]
here is an example of this sort --
[1:26:02 / 35]
I made one, on issues like really challenging material, ie tekstidokumenteille, this analysis, in other words, each word which occurred in the material -- this was on issues like data-mining-related documents, or semmoisia abstract, material -- all the words which of these abstract appeared, was as if entryjä or inputteja data-space .
And the wish was that semmoiset words that appear together a lot, they together define a kind of on issues like category.
So, in this environment could say that together they define issues like generalized keyword, which I hope will be the image of the document in a particular well.
When this multi-word dokumenttiabstraktin material mallitettiin on issues like harvakoodausmenettelyllä, proved that when these features later, then began to look at, as was shown that these harvakoodatut certain features which, for example, the word algorithm was a major concern, and one parallel, one distributed, and in that way, they were very strongly in the present -- or excuse me, was to mention that this scale is now in the picture they are plentiful in 2000 words what was there evidence in the way that they are in alphabetical order, ie the algorithm is here aikalailla at the beginning, and in this piirteessä algorithm was very strongly represented.
So, may be designated for this feature as though it would represent the implementaatioita -- this was semmoinen generalized keyword thereafter.
Here are some of the application, and this is related to methodologies, algorithms, reasoning-related, and some used to relate -- there should be tuommonen human, and collaboration, ie collaboration, and so forth.
[Was there anything in that pre-treatment? ]
Well, harppasin it over --
[1:28:16 / 34]
ie this was done by such a so-called TF-IDF-processing, ie in practice on issues like "the" and "an", they received a very low weight.
This is precisely the description fraktaalinen word frequencies, so their weights were then in this way, and here in middle stages of the weights of words are the greatest, ie those thought to be on issues like specific words.
[1:28:45 / 36]
This is then described, that if we want to do, not only for data analysis, but in fact regression, so far the construction works in two directions, because it neokyberneettisen model structure of this sort is symmetrical.
But this does not speak any more.
[1:29:08 / 37]
Well, these are the association of this degree, that if, for example, are building the animal category, and one of the mouse category, and the horse category, although, as through the mouse and the horse is an animal of this category through the link, then it means that when the horse is active , here also, because of the animal is active, but the mouse is very little active.
On the other hand, if the animal category is particularly active, then the mouse and the horse are reasonably active, however.
In other words, the animal may be characteristics of these mice and horses through the properties.
[1:30:07 / 38]
Well, now could go a little tämmöisiin philosophical questions already here -- here is another Hume and Kant said.
Hume said that the data is not on issues like kausaliteetteja see -- Kant, however, notes that kausalisuus is one of the fundamental features of what the human spirit to build between the observations, ie the person does not see just correlations, but to build kausaliteetteja, ie if a rock flies so the window is broken.
In traditional sense, this is a paradox, because kausaliteettia not really be able to view the data, only correlations.
This is the human mind built this thing kausaliteetti.
Perhaps, however, this kyberneettinen framework provides us with some tools here, as you will recall that the selfish agents' framework -- in other words, this working principle neurons, which were expected to -- so it is based on the fact that instead of a one-way data or information transfer process, so it is these neurons feedback to a lower level, ie, they are eating from the activity below.
So, it sort of what these higher-level neurons of the models, it is at a lower level of these neurons caused by a change in the lower level of activity.
This change in activity is induced in itself came from the outside activity, but in the end, however, neuron models of the u-lines and x-line, namely its impact on the environment there, inducing a data-space.
In other words, a certain way, this build kausaalimalleja, this sort neuron, because it models only its own influence in the environment.
In other words, all the data structures in this sense kausaalisia -- in this certain way, a limited sense kausaalisia, basically at the level -- and in that sense could think that even at senior level, this kausaalisuusajattelu perhaps inherited ago.
So, by implication, I know that what is cause and what effect, because the neurons they know it.
[1:32:58 / 39]
Well, then ekspertiisistä to mention this, that the traditional view is that if x is valid, then one conclusion z is valid.
And if there is another world-power status, x1, then the Z1's conclusion is valid.
If x2 is the so-Z2 is valid.
This will only work if it is just -- knowledge is crispiä mares, or mares exact limit on the mares tradition, just the logical type of 1-0 -- that knowledge is mares either-or type.
But we know that knowledge is really a continuation of mares, that is one of the cluster, a concept that has been elongated this sort -- that is something, a specific concept of a particular material is better and some worse.
Still, a certain conclusion on the basis that it is one of the best category it is responsible.
Now, really, when do we do this reasoning here, as if the best fit in mind, so this is coordinated Gaussiseen model -- if we know that x1 is in force around the world, so when you reconcile this datajakaumaan by the findings, so it is considered that the Z1 is the appropriate conclusion.
So, if there is a continuum of this sort here in x-space, this is also the conclusion of this space must be a continuum, which are the best way possible in this over the distribution sense, then pääteltyjä, these conclusions.
A certain way sumeata reasoning -- but in the way that this blur does not advance to the data of this sort committed feature, as these sumeissa reasoning systems typically have, but it is semmoinen emergoituva also characteristic of this blur, which follows below the structures.
[1:35:20 / 40]
Well, then issues like the knowledge to say this, that when you typically think that knowledge is justified belief in fact, one tune to the concept of the three other concepts -- this really needs to be on issues like hermeneuttiseen circle, will be to define the terms, typically, the three other concepts.
So now on this conceptual definition, is one of the ways to do this sort hermeneuttinen, but because they are not belonging to types 1-0, but they are on issues like the relative, so that different things Tune, that is the truth -- or the tune to the truth, or computer-tune the concept of belief , but the belief tune its part, the concept of information.
And when it is running this sort iterative process, in the way that it sought a balance between these concepts, semmoinen find a balance, a balance hermeneuttinen, without the need to on issues like hermeneuttiseen circle.
So, in a particular way when the quotation marks "truth" is made relative, it becomes universal.
[1:36:48 / 41]
No little wisdom of this sort viisasteleva comment that when it is said that wisdom does not make those mistakes which are intelligent and that then I will survive -- if we say that the smart will be able to cope with problems, so wise to avoid those problems.
This is The point deeper root in the sense that the wise can see the situation in advance, that this leads to difficulties, that is indeed part to avoid those situations, that is not even need to use it for intelligence.
[1:37:30 / 42]
Well, then a traditional problem -- for example, when We talked this Pertti Saari Luoman, so at some stage we always talk about that What about those feelings, that they are not, however, a computer can do.
Well no, of course, can not, emotions are a proper man -- due to the fact that man is a physical object.
But we can kyberneettisessä this framework, a way to restore just the fact that at the lowest level of these neurons take up the activity of what is available.
In other words, if, for example, is the chemical activity on offer, they are gonna learn the information.
For example, if you have one adrenaline levels elevated, and then at the same time, it turns some of the findings of some ghosts, or in some predatory animals, as these are linked together in such a manner that later, these experiences of fear -- they are tied to all these great disservice to a meeting, or we all tämmöisiin hormonal levels, including this level of adrenaline.
They are indeed of this sort very physical dimension to issues like emotions, they are in that sense on issues like low-level emotions, or the lower-level concepts.
And then when we talk about here kvaliasta, so it actually returns to this same case, that all of the findings may have some of this sort of physical dimension also.
[1:39:13 / 43]
Well this too consciousness can suddenly say something, which in this context can be restored -- or that it is not really much to say other than what can fit on a single film.
Traditional thinking has it that the awareness can be defined as that we have this homonculus, which examines the lower-level homonculusta, which examines the lower level, and ultimately it is returned to the lowest level homonculus reviewed by the observation process.
So, we know that we know we know, for example.
So, since this infinite chain of models on issues like the operation is recoverable as if the same level, we can also to this awareness of the problem, perhaps a better bite.
So when we talk about this second order cybernetics chapter as awareness of the problem is that it is äärettömännen single chapter cybernetics -- so all they can be returned to this same level.
In this sense, the animals are in some sense, to some extent, aware of, because they are, they also have a model on its activities.
And a real awareness comes a stage when the upper-level model is the mallituksen level that it detects that it is reasonable to form a model in which the environment and your own self are separated elements -- they formed their own on issues like agent mallittamaan them.
A way, the stage when it comes to their own awareness of their own, or I, the self-performance model, so it can be said that the system is aware of.
At least these are the semmoisia definitions of what is to be used in the definition of awareness, but that it has very much been a tangible benefit of the tradition.
[1:41:23 / 44]
Well, then this knowledge can, of course, continue up, so if we understand what others think, seamlessly, so we can form a higher level of consciousness, or something of this sort systeemiäly, or intelligent level organizations.
Those issues have now then just their own courses here.
[1:41:51 / 45]
This should really be quite a lot of things, but we have to break off at some point.
But it may be concluded, however, that it that it is indeed a range of levels of these, now harvakoodautuneita these elements, so it brings some kind of added value to this information processing.
If you consider that all the information that comes in with sensors, it is logarithmic -- for example, visual data, the intensities are given in logarithmic, kuuloaistimus is logarithmic.
Then it can be argued that these findings, the mean value of logarithms of the sum is the same as the original input signal.
A certain way now tämmöisten combined sum of the form is the entry form, combining alkuperäisisille signals -- that is to talk about issues like the AND operations, if they are sufficiently independent of each other they inputit.
Well, kyberneettinen model to build the most rippumattomia, or those applying for independent components, that maybe this can be expected.
Well what does this harvakoodautuminen do?
It means that in the alternative, either these or these interpretations, it is this sort XOR-type konnektiivi way.
So, this is this sort AND-XOR graph what the neurons of this sort a chain of these logarithmic signals does.
[1:43:29 / 46]
[1:43:31 / 47]
In other words, practice makes mares miks luck mixture models, and it can start to explain the fact that how a person might, for example, finds a table.
At the lowest level are the pixels correlations -- they will be to interpret the elements of line, or else rods and so on -- and when there are four pieces on issues like rods, and at the same time occurs in a disk of this sort, so it has some relevance, or a certain credibility that it is already in a table.
This comes back to the environment feedback, that if there is a chair, or if there is still a higher level of understanding that this is a classroom example, or a meeting room, then there will be upper lisärelevanssia of this Protocol the interpretation of -- and ultimately when you have achieved this balance between these concepts, the interpretation of the way to assume that this interpretation of the Protocol has won all the alternative interpretations, that is, XOR structure is stuck in a particular interpretation.
And this may be able to use even a kind of language, on issues like a deep structure of a model.
[1:44:43 / 48]
[1:44:44 / 49]
[1:44:45 / 50]
[1:44:50 / 51]
Well, here kyberneettisestä semantics, I would like some to mention.
Because when you are building on issues like multi-level models, so of course, can start to think about how that can be guaranteed by the fact that we have all the relevant information included in the data that we can built a meaningful model, so if it comes to people, even though the model, so -- with the man -- so, how much keep them inputteja take in order to have a credible on issues like attraktorit time.
That way, in this limited space is found attraktorit same as what I find, when it is there for a larger space.
Well, it can be assumed that for this kyberneettinen machinery would be able to mallittamaan something at a lower level kyberneettisiä systems -- in other words, this chain kyberneettisiä these neurons form a lower-level model kyberneettisestä model, and then to form a kyberneettiselle model, the model must in some way it is an integral part of an essential or thing, ie the balance there, among the neurons, to be able mallittamaan.
Well, can not be mere data to see the balance, but instead, in the case of these forces are involved, the forces that are driving forward the change in the system -- ie the sort of issues like years to mallittamaan, then it can also be in balance next mallittamaan flow.
And if this can really flow mallittamaan, so can mallittamaan also those forces capable of this flow to prevent, namely the forces.
In other words, if these forces in a given situation may be mallitetuiksi, then it will balance mallittamaan.
This review is a meaningful sense, that we get here a little higher level in the fact that what really want to model.
[1:47:10 / 52]
After all, the models to be on issues like gradient there, how the neurons change -- meaning that they reflect the flow of mares that what is going.
We expanded this really a chess game example in this way, that when it is an example of what we discussed it earlier in a lecture at the beginning, so it was only mares static situations mallitusta -- it does not in any way to take a position on how that would play.
In chess, is precisely the problem interesting is that the consensus, or expert consensus, to find the hot points, on issues like hot spots, in this case, a chess board, that what happens to interest -- so they are precisely those points at which the chess game flow, that is, a game a year, passes -- if it is seen in similar situations, so it is the direction in which it flows strongest in the different scenarios, just as it describes the flow of the interesting points, which in this case, it could be described as gradient.
Now, when you mallitettiin the chess situation in addition to this the situation and the situation of the difference, so we saw that it really started to model these hot spots, which means it started to claim that all right, this situation is likely to move the queen of that box to that box -- which is quite reasonable in itself.
But the truth was that it was the tower that was moved to another location.
This gradient was carried out exactly in the way that two successive binary presentation was reduced to one another, when shakkilaudalle may become a negative inputteja, or issues like the minus sign the pieces -- that they are negative pieces in this present, with those in the wrong pieces.
[1:49:23 / 53]
Well, then these gradienttipiirteitä they can be used to start a certain way on issues like to describe the projections of the credibility of space -- but this harppaan over.
[1:49:36 / 54]
And a certain way if ruvetaankin mallittamaan successive issues like gradient, so when it comes to adjust the gradient of these away, so it comes to making mares -- on issues like changes to try to adjust out.
So, the long term, all the highest derivative nolliksi go, that is intended to alter the world through a static, or in practice, thermodynamic equilibrium.
So back to the last time on.
[1:50:18 / 55]
Well, it is when -- harppaan here -- that when it comes to these gradient mallittamaan, then realize that this is a very big problem, in practice, the change around the world -- we have to practice to predict the future state of order at this time we could of this state gradient models, and it follows the fact that higher level, this implicit control what these neurons are doing, so it becomes explicit säädöksi,
[1:50:53 / 56]
for example, what this cat does this when it takes the mouse off, or a dinosaur when it translates to a head in the observation direction of change in visual field.
[1:51:04 / 57]
It follows from the increasingly higher-level forecasting mares, and that on issues like planning and mares up to the imagination then, that is it comes in its own world, to build a variety of alternative scenarios.
All of this supports the fact that we are able to build better models, or to provide a better environment.
Here, at a higher level, it no longer seems to be no problem adjusting.
[1:51:35 / 58]
Well, then return to the next time that the end of the day, however, such data filtering, after all, is it a credible model of awareness, when, after all, the most interesting things are just on issues like the imagination relate to, or issues like wild Associations -- just different things at once, they pystytäänkin to connect to each other , as if to make on issues like tree-transformation neuraalisten between the structures.
If you want to really believe in the way of modeling cognition, we must somehow be able to respond to tämmöisiin, or give any idea of the direction -- and next time look it.
Then here is another thing, this, that what can be said about issues like,
[1:52:26 / 59]
on issues like -- then that how this assosiatiivista, how there in the activities can be discharged back to a linguistic -- it is a problem which is now actually possible in this context, or does not even have time to address.
[1:52:47 / 60]
But we note, however, that the very same problems are here in the cognitive and the biological side,
[1:52:53 / 61]
and maybe at some point it can be said that a certain way the same methods are appropriate to issues like a biological or genetic information processing, than what we have now is the understanding of cognitive systems.
Well, this is only on issues like heittona, and a reference to the fact
[1:53:13 / 62]
on issues like that analogy may be of value -- that is, next time look analogioita.
Well, thank you.
... once again, went even further extension.
[1:53:32 / -]
(v.2009.04.10, only a rough machine translation, not cleaned yet!)