[deespona]

What if we have no memory at all? I mean, all Hawkins method is based in seeing the brain like a memory system, that records and stores billions of pattern signals from events of the past. Also our AI method for computers (TUE tridimensional understanding engines) is based mainly in a memory system, as it works with tridimensional forms that allows the system to construct 3D scenes, then afterwards the system decides actions to be done with the TUE. This means our system is really a mixed one, as it first extracts from a memory ( a 3D library ) and after process actions, reconfiguring itself to match different 3D puzzle pieces to obtain a solution.
But maybe the brain can be pure process, that is, the brain maybe don´t store a single issue, but configure progressively a "refinement machine", that according to the inputs is capable to ellaborate a system of conections that allow us "rendering" something than looks us similar to memory. We don´t remember things from the past in a single step, but we reconstruct events in a very aproximative and progressively detailed and inexact, not at all "photographic" way. The idea is that our brain maybe don´t works like a DVD, but more similar to a videogame, making each image or memory event not extracted from a file but rendered in real time, different for each situation. On despite in the videogame exist previously memorized 3D objects and textures, like in our TUE system, in our brains also these objects and textures can be perfectly a more "primary level" result of simple, maybe never exactly repeated connections, not obtained from a "data cell" or a similar "array based thing" containing a pattern record. This is totally opposed to the Hawkins memory system, but after a full year in this forum I think a "pure process" brain model must not be still discarded, on despite the fact that all the inputs are only patterns.
Jose Maria de Espona
http://www.deespona.com/
http://www.inteligencia-artificial.com/

[chatham]

[deespona]

Our tridimensional method is only related to machines. If a similar priciple works or not in human brains is secondary to us, as this is not our area of research. As you can check in our web site, the method is named tridimensional or tetradimensional, according to the type of TUE involved, if time is considered or not.
Our method for computers uses something similar to pattern arrays only in a phase of the proceeding, specifically to recon 3d objects in a virtual space by generating ortonormal profiles of the 3d object, then generating patterns to be compared with other signatures in a library. It´s a method developed in 2002 by a computer graphics department at Priceton, and it was developed using our 3D library as the basic reference. Not all 3d objects recognition methods are necessarily based in pattern comparisons, but we use it slightly modified because is very fast and reasonably accurate, and well, we are familiar with it.

But my question is more dramatic than using or not the Hawkins memory framework. My highly speculative, intuitive question is related to the posibility that the brain really doesn´t store any type of information inside, but maybe consist more in a "filter", in an auto-organized pure processor that "renders" images, words or any memory data-like issues, that are trigged by the inputs, but not necessarily store these inputs or output results inside the brain, a sytem that only is focused in creating the "neuron paths" to give some "exit" or "solution" to the input pattern arrays.
Sounds hard, but in a system like this one, the equivalent to the memory library would be the reality, the reference system OUTSIDE the brain.

[Lawrence Phillia]

[chatham]

[deespona]

First of all, one important question: since 24 hours ago I can´t enter in this forum from my old IP address, then if the webmaster of the forum reads this one, please check if it has been blocked accidentally, and allow me to enter again here from my old IP. It´s a bit embarrasing to use the IP and PC of a neighbour company, here in Madrid.
I will try to show here how this question about a "no memory at all" model of the brain has arisen. The cause has been to discover a paradox in our TUE system, then as I am not a biologist (on despite my wife is a viral biologist, my level of knowledge about brain is few more ellaborated )than those publically available at Sesamo Street show) I was just curious to check with top neurobilogist, like those present in this forum, if this "no memory at all" possibility can match the brain´s detected performance in some way.
Our TUE system is composed of two main pieces: First a 3D library, that contains 3D objects that exist on the reality, encapsulated in a file with the 3D form of the object, a set of 2D profiles to identify the object, some related names to identify the object in different languages, a set of tags (handles, action launchers, etc..) associated to the object or parts of the 3D object, and physical properties (probable weight, size, color, noise, termic profile, magnetism, dynamic performance, etc).
Second, the TUE, the tridimensional Understanding Engine associated to an arbitrary object (A cube containing some letters) that we say is the computer itself. The TUE is basically a scene renderer, that is, after receiving any type and number of parallel stimulus via any way (words entered by microphone or keyboard, images via videocamera...), the TUE generates secuentially a 3D scene containing the most related issues to the input stimulus. The TUE has also an "understanding ratio", the number of input stimulus that can accept to be operative and not be blocked by stimulus saturation. The TUE works like "I see the characteristics of these objects in my scene, then I will search for ways to match all of them together in a coherent way, that is, rendering a solution that can be another image, another word or inclusively another stimulus for continuing the process.
The TUE actually reads the objects from a library, because it has been designed in this way, but what if the external stimulus contains inside the same properties (3D form, physical characteristics, etc) that are present in the defined objects of our 3D library? What if the TUE founds by itself a WAY to "read" the input stimulus, then rendering solutions without the need of a 3D library?. The equivalent in an organic brain would be a system that accept any type of pattern stimulus, then create a wave of conections to render a scene. In other words, our memory area is maybe a type of analogic, sequential "rendering engine". No "memory library" is necessary at all.
Obviously the TUE needs also learning: re-inforcing ways for seeing the diferent input stimulus. If you have more input stimulus readers you are more intelligent, because you can detect more and more properties from the stimulating object or issue. If additionally the TUE does it fast, if you render a reasonably good solution in less time, you not only are intelligent, you are also smart. If this is not enough clear because my english is not very good, I will try to give you some cotidian examples. And I hope to do it from my old IP address then!.

[deespona]

Thank you!

[deespona]

As some readers of this forum have decided to use my personal email for giving me their not very kind opinions about my ideas, endorsing me to return to my Sesamo Street Home, then I feel pressed to send another open letter to the forum.
I also think Jeff Hawkins is extremelly important for humankind, not only for his book and researches, but also, and this is not a banal question, for creating a serious and multidisciplinar open forum like this one.
For this reason I think about the question, and I shall try to formulate a compatible model between an hipothetical "no memory at all" model and the Hawkins "memory prediction model". I believe this new reductionist model will satisfy many opinions, and also I will try to no mention my empirical, commercial, and auto-promoted TUE method. I will try to do it in your neurobiologistical language code, on despite I can do monumental errors. We will start with this megalomaniac sentence:

"We call memory to a specific electrical path created across a neuron network, that modulates an input pattern signal and returns at the end of the path a new, modulated pattern"

In this model the input pattern can be related to, by example, a sound, but the output pattern, by its form, can be more related to an image. Here the output new pattern can launch a new wave of electrical signals to obtain sucessive new patterns, in a progressive refinement process. If an used neuron path is revisited, then this path is chemically incited to be phisically reinforced, by example with a bigger diameter at the axon conections, to guarantee a better synapse. Also near, parallelly used paths can influx electromagnetically bettween them, forcing the dendrite grow conections between correlated, used paths.
The Sesamo Street boy can also imagine in his colorful world that this maybe can explain why is necessary to sleep. During our "vigilia" status our brain is under a continuous electrical storm, then we can make associations, electromagnetical discharges between different not connected electrical paths, but these temporal electromagnetical bridges will launch chemicals footprints at dentrite´s ends, to consolidate a new, physical axon based bridge when the electrical storm will be reduced in intensity.
At the top of my delirant fantasy as Sesamo Street official neurobilogist "dilletante", I can adventure that "dreams" are probably new axon-bridge testers, after the biological grow of the connection is finished during our sleeping process . In this model, nightmares can be the result of aberrations in the night bridge construction, creating finally wrong patterns that forces us to awake, ending in this way the physical consolidation of an erroneus synaptic connection.

Here is, a concilliating theory of "memory prediction framework" and "pure processor brain" model.

As many of you have cats, monkeys and rats supceptible to be electrocuted, pls check if this theory have some experimental base. Only for fun. But please don´t tell me that these pets suffer! Remember, I live in Sesamo Street.
Sorry, no more emoticons.

[the_pm]

(Posted by the_pm on behalf of Pankaj)

I am actually very impressed with your way of thinking. It is out of the box idea. To crack the problem of how brain works, we need a breakthrough. Nobody has come even close to explaining it. My gut feeling is that it is going to be a relatively simple solution,(I mean without invloving
Bose-Einstein theories) but I could be wrong!

I am in a totally different field, but after reading On Intelligence, my childhood curiosity about brain has been rekindled to the extent that I just keep thinking about it most of the time. I have no knowledge of neuroscience, but have read bits and pieces. About three weeks ago, I realized that image recognition that everyone is trying to crack, using
current algorithms using neural networkig or Bayesian network will not be possible. In my views, the key to that lies in what you have earlier proposed (TUE? I have to read more about it). The key to recognition is
learning in the 3D model, and then projecting it to 2D (image). When you look at a picture, you know what is background and what is foreground based on your 3D model. Learning is done by not just a static 3D, but a dynamic one, which is also dependent on time. Thats not all, its actually much more complex than that, as during learning, its not just the visual pattern, its orientation has to be involved too. You can not recognize a face upside down as fast as you can a straight face. When you look at a picture of a person, and then look at another picture of the same person in a different pose, the regular algorithm will fail if there is a broad smile. This is not an error in the picture, but brain knows the same person is smiling. One thing is for sure, that the brain does not store all orientations of every object. It would be very inefficient. I am still debating about whether it stores a pattern, or what you suggested that it doesnt store any pattern at all (?). It does identify a pattern, and does remember it and applies it to similar situations. An example of it would be if you throw a ball, you can catch it based on your learning, but you can also catch a block of wood, as its projected path in air would be about the same as that of ball. If brain stored all patterns, it still would be a huge amount of informatio. About two weeks ago, when I was thinking harder about this, it occured to me that may be brain doesnt store exact information at all!! Brain has about 100 billion cells (Neurons). Storing a picture (as we know) of every thing you know is impossible in that. When I read your thread today, I was just excited that there is someone else who thinks that way! The brain does store information though.

People (your critics too) often try to disparage new ideas as -it doesnt explain consciousnes. We should try to crack the code one step
at a time. The more I think about brain's recognition power, the more I realize that the most amazing thing to me is that we are able to see the world as it exists- We go outside and see the road, tree, stones, cars etc
and are able to see as they exist. We are able to tell what objects are together and what are separate. We can distinguish that two stones together are two or just one stone in a weird shape. I get more amazed
that I can accurately see a dent in a car and less by recognizing
which model car it is. (Image analyzer will probably think it as statistical error in pixels?) I can go on and on, I hope you get the point. May be you are using same model in your TUE system. I think recognizing the worldly objects will be the point to start learning how brain works. This should be the easiest(!) of all problems, as I think most animals can recognize the 3D model of their world. Also the fact that babies are able to see and have a 3D model of it much sooner than they learn to read/write/talk. I also think that once we learn the basic model, we can then progress to add corelation between them and start towards more intelligent ways.
I think the process will be the same, as humans just have more cortical cells than animals, that can take identifying patterns to the next levels.

Best Regards,

Pankaj

[deespona]

Pankaj, thanks a lot for your support and for sharing with us your comments. I sincerelly wish that also the hard critics of my proposed model of the brain will publish here their comments, and not only in my personal email box. I´m sure Jeff will be more pleased in seeing his forum is really useful to post more interesting models and simulation topics about intelligence, freely and openly debated.
Like you, I think more and more solidly that a 3D interpretation and approach to the problem is the key issue, and the solution not only for constructing an operative and stable AI method, but also to explain a plausible and coherent model of the brain.
As the time is running, we here in my company see more and more satisfactory implications of a model of the brain like a "filter buffer-analogic modulator" as a result of the surrounding tridimensional reality. And we are convinced of this "no-data-at-all-inside brain" model because on despite is a deductive, non experimental model, it fits perfectly with our equivalent computer based system, the TUE, that to a considerable extent is an empirical and working solution.
By example, take the short and long term memory problem. If the brain is a constructed and progressivelly refined group of physical and electromagnetical intra-neuronal paths for modulating input patterns, product of the external 3D world, or from other just-inside processed patterns, then to explain how they can work in the proposed model is very easy: short term memory (those we can compare with the Hawkins´associative memory) can be the result of real time, combined physical and electromagnetical paths between neurons under input pattern stimulus wave storms. The huge majority of these temporal, associative electromagnetic paths will not survive the sleeping process, as these electomagnetic mini-discharges between neurons will not have the enough electrical intensity to initiate the chemical reaction that stimulates the dendrite grow to another neuron, then to consolidate physically a new modulating path for electrical signals. Higher the pattern stimulus, higher the electric dischrge, then higher also the chemical discharge and the consecutive growth of the physical dendrite-to-axon connection. In my model, obviously the axon connected, stable physical paths are the long term memory equivalent.
And this can explain more: old people can´t consolidate short term memory issues because, with older age, they can´t produce so efficiently the chemical reactions for provoking axon and dendrite grow, on despite of relatively huge electromagnetical discharges. Only the very huge and old axon paths of the infancy, so many times activated in the course of life will remain, generating more and more self-induced pattern filtering, rendering the most old remembered issues and traumas of the infancy to the last moment, like the dramatic last scene of Citizen Kane, with Orson Wells remembering his snow slide, traumatic symbol of being removed from his youth´s environment. With time, also the most huge axons will fade, not allowing more external or self-induced generated pattern processing.
Jeff, if you are reading this one ( but I doubt Jeff really exists, as I have not found a single post from him here, and I doubt on despite I have seen him two weeks ago in an TV interview in the spanish public television, speaking with Edouard Punset) say I hope you consider my hipothesis of the brain enough solid to call it "The 13th Prediction". And if this extravagant but coherent theory is confirmed experimentally, please at least allow me to call it the "Hawkins-deespona" model of the brain. Or, better and more armonious sounds, then it will be better remembered in a neuron path, the "Deespona-Hawkins" brain model. In any case I will tell this story in this way to my grandsons, when available.

Now seriously (well, a bit more seriously), if a model of the brain like this one is finally confirmed experimentally, the consequences can be extremelly dramatic in many fields. One of the incredible and fantastic consequences of this brain´s model discussed here is that this "Deespona-Hawkins" brain model can leads in some way to immortality. Yes, we have arrived here to this technical, non scientific but "technical state-of -the art" current possibility. But I prefer to discuss this question in a new topic, if the webmaster are on disposal to allow me to do such crazy thing.

[boundlesslife]

[pcerkez]

[boundlesslife]

[deespona]

Er...After one year out of this forum as a result of the *misterious* web site down ocurred one year ago, I return here with the same essential questions, concerning the issue of "no memory at all" model of the brain, that is, that the brain is basically an "anologic filter" of the reality, converting input patterns into output patterns through neural connection nets developed during our lives. Based in this principle, the "memory" is really in our environment, and the brain simply transform input patterns that comes from environmental captures by our senses into reaction patterns that allows us to interact with the environment. No data stored at all.
As maybe few of you remembers, we have developed an understanding system for robots based in an implementation of the idea, the 3DFORM-ID principle, where all the data is transmitted by the environment to the robot, then the robot is only a "filter" of the reality, and doesn´t store a single bit from around, but simply process it to generate an action response to be done in this environment. In this model the meaning of a concept (by example, a "chair") is the shape of the object related to the concept (the shape of the "chair"), then the robot can "understand" the "chair" because it can filter the 3D data dowloaded from the object to create a response in the form of an electrical pattern (move to avoid the "chair").
The implementation of the idea is drastically simple, and we are just in the prototype phase, by creating a device named "3dturtle" attached to a physical object that contains 1. The 3D shape description of the attached object. 2.The 3D position of the device into the object´s 3D shape description, 3. a spatial reference triangle into the 3D shape description, relative to 3 spatial pinpoints or tags that can be detected by an external system, and 4. Other physical properties and program code to allow a correct manipulation or response by the robot. This last is important, because this means that the robot practically have no stored data at all(memory) but a very simple 3D scene creation proceeding (the equivalent to our neural brain net of axons based paths), and have no programs at all inside the robot to allow the manipulation of objects into the environment, on exception than those that comes dowloaded from the own 3D environment.
To those interested, we have created a site with more info about these operative "3DTurtles" at http://www.3dturtles.com

Jose Maria de Espona

[intelligent robot]

I see little difference between your ideas and Hawkins' framework.

He doesn't assert that we store everything in memory; he just implies that we store important, critical information (what he calls 'invariant representations') in memory, then use our hierarchical circuitry to apply that to a wide variety of situations.

If you think your idea is very different from this, then please elaborate on it.