New brain/mind theory

My website “Quad Nets: Device Models of Brains” is online at http://www.quadnets.com.
(link)

“Quad Nets” proposes a new kind of “artificial intelligence” that uses devices other than computers. Chiefly presented through Images, the Quad Net approach integrates physics, neuroscience and psychology in primal forms, initially rudimentary, but suitable for unlimited development in size and complexity.

I am an amateur and have privately worked in these areas for many years. Unfortunately, I have not found a means of communication through established channels. I hope that the readers of this blog will provide needed critical review. Thanks to the “neurodudes” for making this medium available.

Bob Kovsky (rlk “at” sonic.net)

Advertisements

17 thoughts on “New brain/mind theory

  1. Utter BS. Poorly written paper. I can’t say I read the whole thing, but I sure couldn’t find a definition of what the modular units do. No connections are made either with experimental data beyond vague qualitatives (oscillation! just like EEG! oh, stop the presses), and the “networks” or whatever they are, don’t seem to solve any problems other than creating garish animated gifs.

    Like

  2. Bob: it is very hard to tell, from what you’ve said here and at your website, what a quadnet actually is. It would really help if you dissected an individual ‘tial’ mathematically. That is, what rules do they follow? What are the rules for how they interact with one another? Is there an example of a computation that they can perform? E.g., xor. Can the connections be altered with experience (i.e., learning)?

    Like

  3. Hi Bob,

    Here’s my advice, from 20 years in the industry.

    1. Define what each unit does.
    2. Describe how the units interact.
    3. Show that the interaction causes the units to generate a representation.

    Like

  4. I wrote “Quad Nets.” Thanks to you who looked at the paper and commented.

    My overall response is that “Quad Nets” is a new approach and requires a reader to be receptive.

    The important activity of a Quad Net is collective, not reducible to units. There is a repertoire (set) of activities (see Image 11) and you can’t tell which activity is ongoing by looking at a single unit. Elemental devices are organized into a collective, making up a device (e.g., a TQN). In operation, the entire Quad Net, initially silent, enters into activity as a collective. The idea is like that of a physical material that changes its entire mode of behavior collectively — like ice changes into water. (Details are presented in Section 3, where a selection is analyzed in ways the model allows.)

    Of course, as the comments emphasize, a presentation must include definitions of activity of individual elemental devices. Units (elemental devices) are thoroughly defined, along with interactions, in Section 4. The definition is very simple — complexities come with collective action. The beginning point is the “Primal Pulser” (Image 2). All that happens is that a unit pulses with a characteristic period.

    An elemental device is like a container with an off-center pivot that constantly receives water and hold the water until a certain “critical point,” when it tips over, spilling all the water, then returns to the original position, to start again. Fancifully, if containers are arranged in an array, the spilled water from one container adds to the water in other containers and they tip over in a wave. The water metaphor is weak because water only falls down, while Quad Net interactions are effective in all directions. Note that Quad Nets are not finite state machines and this distinction makes the concepts more difficult. [My conjecture is that as device operations approach the Critical State, computational emulations of Images 35 and 54 “blow up” by requiring finer and finer time periods for convergence. Each individual device part has its own clock and periods of clock ticks are separately variable.]

    Using the water metaphor, a variable flow of water into a (collective) device controls both the timing of the collective spills in that device and also how the collective spills in one device affect other devices. Metaphorically, variations in water flow in different device are coordinated to govern activity.

    The approach is rudimentary and demonstrations are by construction. The idea is similar to that of computers, where you start with a conceptual system and look for applications. I am not trying to perform computations or to model learning. [These goals are in view, although far away. E.g., Images 33 and 34 model a “Clear and Add” computer instruction.] I discuss an imitation of jellyfish (“push, push, push) and the next step in this line is to imitate a sponge (“stretch, squash, stretch, squash, stretch, squash).

    A simple engineered organism is a waterborne “GoUpper” that has four “flippers” in a square arrangement and that adjusts the twitch rate of the flippers to maintain an upward direction. The adjustments are controlled by a selection device based on Image 38 that contains an “otolith” like those in the inner ears of animals. During each cycle, the variable position of the otolith in the selection device sets the variable flipper rates for that cycle.

    In another line of development are “peristalsis” designs where muscular units and controlling Quad Nets are arranged in cylinders around a compressible tube. A Quad Net drives waves of constrictive muscular activity down the tube, pushing material stuff (a “bolus”). Another Quad Net with hairs projecting from each elemental device detects stuff in the tube and directs the muscular activity so as to relax muscles ahead of the stuff and facilitate the pushing.

    The focus in “Quad Nets” is an integrated approach to physics, neuroscience and psychology. The approach is by construction and development from the bottom up rather than a global theory from which results are derived.

    I hope that critical readers will give the website another chance. There is an overall survey through the Images that is intended to motivate the more detailed formal paper. I believe that new ideas are needed in these sciences and, whether or not mine have merit, acquisition of new ideas requires time and willingness. Your further interest and comments are very welcome.

    Like

  5. For those interested, the main paper is here.

    Bob: I don’t really understand the paper, but that may be because it draws on concepts from materials science and statistical mechanics with which I have little to no competence. If you want to make it understandable (and, hence, read) by your average neuroscientist, I think a more basic presentation would be needed. Instead of just making allusions to concepts from fields foreign to most of us, you should probably be more explicit, especially providing some equations in the simplest possible cases (e.g., one tile), and clearly analyzing the simple case so that a reasonable and intelligent person could understand it.

    I would suggest trying to present some of your ideas at a conference where people who know the relevant fields will be in attendance.

    Like

  6. Thanks to Steve for his comments.

    Quad Nets is on the “borderland” of three disciplines: 1. artificial intelligence – by which I mean imitating activity of intelligence with devices; 2. neuroscience; and 3. physics and related physical sciences. As to physical sciences, Quad Nets are novel and rather speculative extensions of principles of statistical mechanics — the forms of content in Quad Net constructions are activity patterns while the forms of content in statistical mechanics are momentary states. I justify the extension by invoking the universal nature of Critical Point principles and other principles discussed in § 5 of the paper. Regardless of whether the justification has merit, the ideas are novel and non-standard in the physical sciences (as they are here).

    One of the chief features of Quad Net constructions is collective activity, requiring multiple elemental devices. Activities of an isolated device are shown (Images 40, 45-46) but these activities are not very interesting. The nature of collective activity changes as the number of elemental devices in the collection increases. The smallest possible number is 3 (needed to support a cycle) but I expanded this to 6 to show some variety in the repertoire of the Simple Cycler. (Image 39 and Section 4.a of the paper.) The formal presentation is by way of Images not by equations. Images 44 and 48 show exemplary low-level and high-level activity patterns maintained in the Simple Cycler. Primal Quad Net (Images 7 et seq) has (effectively) an infinite number of elemental devices. Toroidal Quad Net devices have intermediate numbers, e.g., 400 elemental devices in the TQN’s shown in operating designs in Images 32-35.

    I am sincerely grateful for the opportunity to discuss my work. I think my best hope is an online forum such as this one. This is the first place I am trying to publicize the website. Perhaps someday, someone will conclude that these discussions had substance (ROFL).

    bob kovsky (rlk “at” sonic.net)

    Like

  7. I am reminded about many of the new ideas that emerged that allowed for heart operations… and how many threw the new findings aside. Others… like you though … can in with an open mind. The findings are rearely perfect at first … and some even suffer because of their imperfections (heart transplant patients all ded in the first few tries) but discovery is what keeps many of us alive and hopeful. Thanks for this one — I plan to look into it even more because of your post!

    Like

  8. Bob: without an equation to describe how an individual element works, it’s impossible for me to say anything more. Even if you need six to get a cycle, it would be very helpful to know what is cycling, what the rules governing the individual elements of the cycle. It is usually pedagogically most useful to start simple, with equations that govern how an individual element (whether it be neuron, capacitor, etc) works, and then build up slowly to more interesting behavior. Word models can only take us so far.

    Like

  9. Bob Kovsky responding to Eric Thomson’s most recent post.

    Thanks for continuing the dialogue. Your posts (and that of Mike S) make clear a preference for an “atomic-molecular” approach where, first, elementary units are fully defined, then elementary units are hooked together with simple rules and, finally, desired activities occur as a consequence of the systemic organization of the units. This is a traditional approach that is used in mechanics (adding forces), chemistry, propositional logic and computers. I can structure such an approach for Quad Nets — e.g., using Images 40 and 46 for the activity of a single, isolated elemental device. These Images have the form of graphs and could be expressed, e.g., as sawtooth functions. However, my attempts to express the activity of a Quad Net in a mathematical formulation show no benefit. More, this approach misses a major point of the Quad Net model, namely that the different forms of activity are collective and, sometimes, you can’t tell just by looking at one elemental device which of several phases is being maintained. My approach, apparently imperfectly expressed, is to identify the collective activities as primal and individual activities as representative. This presents the material in a different order from that of an atomic-molecular approach. Hence, the “atomic” models appear in section 4 and the collective constructions in sections 2 and 3

    Like

  10. Bob said:
    More, this approach misses a major point of the Quad Net model, namely that the different forms of activity are collective and, sometimes, you can’t tell just by looking at one elemental device which of several phases is being maintained.

    The same can be said of cellular automata, but the rules governing the behavior of individual cells is quite explicable. My main point is, that as a sociological fact, few people will read or understand your theory unless you take what you call the ‘atomic-molecular’ approach. Nobody will make it to section 4, and words/images are just not enough. Take a look at any introductory book on artificial neural networks to see a better way. (E.g., Neural Network Design, by Hagan).

    Like

  11. Bob to Eric: I appreciate your experience with “sociological facts.” In effect, I hear a suggestion for an alternative ordering of materials in the paper such that section 4 appears first, to be followed by section 2, section 3, section 5 and, finally, section 1, with suitable revisions. The new section 1 (old section 4) would first define an activated elemental device – like the Primal Pulser – that goes beep-beep-beep with timings controlled through interactions from other elemental devices and/or the engineer. (This contrasts with traditional neural network models that start with the Hopfield Model or a McCulloch-Pitts neuron, as in Intro to the Theory of Neural Computation (Santa Fe) by Hertz, Krogh and Palmer, physicists who develop a statistical mechanics approach.) Then elemental devices so defined could be assembled into a 3-cycler (3 phases in the repertoire), a 4-cycler (same), a 5-cycler (5 phases), a 6-cycler (the Simple Cycler, same), a 7-cycler (7 phases), etc. Large-scale constructions in subsequent sections would be based on the opening material. This would be an alternative presentation of the materials and would undoubtedly produce some new constructions.

    Comparing the proposed alternative with the actual paper, it is the chicken-egg question. Which comes first, the organism acting purposefully or an individual elemental device producing shifting patterns of pulses? The question is especially difficult in the alternative proposal, where it is necessary to navigate a transition from individual time patterns in elemental devices to collective space patterns in Quad Net constructions. Supposing you can build a chicken up from an egg, sometimes it is better to start with the chicken and focus on the egg-producing system.

    The design of the paper is grounded in the foundational concept of tiling. In Quad Nets, tiling is universal, primal and ideal and occurs in space, time and activity. The paper shows many and various tiled constructions. The purpose of elemental devices is to participate collectively in a tiled construction, not to model a neuron or to conform to a technology like computers. A tiled construction has character, however tediously reproduced. As stated at the outset of section 2, we are not in a mechanical system that starts with an empty void into which atom-like devices are introduced and assembled into systems according to a molecular principle. Rather, we are in the interior of something like a physical material with a particular collective character that is an actual starting point for what is going on. The details of a graph in section 4 are only secondary and other graphs would produce the same results if used in Quad Net constructions.

    The present organization of the paper is also based on an approach that emphasizes biological and neuroscience models above engineering models. Quad Net constructions are “mesoscopic,” using a word that appears in a title of a book by Prof. Walter J. Freeman (link to online book on the quadnets website) and that denotes something in the middle between macroscopic organism models (e.g., stimulus-response) and microscopic neuronal models. In addition, an image of Quad Net material being fabricated during extrusion in a manufacturing plant suggests an answer to the question posed in Edelman’s Neural Darwinism, p. 75, “How does a one-dimensional genetic code specify a three-dimensional animal” and the suggested Quad Net answer resonates with the biological answer Edelman provides. It is reasonable to suppose that the path of Quad Net constructions resembles growth of brains in biological environments, with initial crude constructions providing points of attachment for later, more refined constructions. It is not reasonable to suppose that building a computer (or a neural network system) looks like growing a brain.

    Like

  12. Bob wrote:
    In effect, I hear a suggestion for an alternative ordering of materials in the paper such that section 4 appears first, to be followed by section 2, section 3, section 5 and, finally, section 1, with suitable revisions.

    Exactly! But also it would be good to fill out in more detail the input-output transform being implemented by an individual element: a drawing that specifies the inputs, outputs, and the transform between the two. In particular, does it take in real numbers as inputs and output real numbers? That kind of thing. For block diagrams, the systems people first clearly delineate what each block is doing (e.g., addition, multiplication), and then analyze how these blocks work when they are all hooked together.

    Like

  13. Eric wrote: “it would be good to fill out in more detail the input-output transform being implemented by an individual element: a drawing that specifies the inputs, outputs, and the transform between the two. In particular, does it take in real numbers as inputs and output real numbers? That kind of thing. For block diagrams, the systems people first clearly delineate what each block is doing (e.g., addition, multiplication), and then analyze how these blocks work when they are all hooked together.”

    Bob respondes: Well, let’s say a Quad Net device (or a neuron) is a like a Krispy Kreme factory that produces doughnuts (or beeps or action potentials). Different kinds of stuff are inputs and the list of possible ingredients is open-ended. The controls are more important. Learn enough about the controls to get a production line going with any convenient ingredients and then you can work out the variations. Inputs are non-specific to the question of function.

    The output of any Quad Net device is the same — it goes “beep-beep-beep” — but, in large devices, the beeps are organized into patterns in space and time. It’s the patterns that are important, not the individual beeper. Patterns go in and patterns come out and patterns drive the going ins and the coming outs. There is no “transfer function” – changes in beep patterns are based on in inputs and controls and depend on particular configurations of parts and particular operating values of variables. There is baseline operation that can be changed by various means and the particular means depend on the circumstances.

    Quad Nets are the mother of all “block diagrams.” Quad Net constructions are working block diagrams. A Quad Net spatial block (e.g., a Flat Quad Net shaped into a CQN or TQN) is activated and the activated block turns into a device. A Quad Net device embodies a working block in a block diagram. Hookups between QN devices embody lines in block diagrams. Each device transfers patterns to other devices and/or controls transfers between devices. Devices are hooked together into tiled assemblies; and, further, tiled assemblies are hooked together into nested tiled assemblies; and then nested tiled assemblies are hooked to each other. There are also hookups between devices in one tiled assembly and the entirety of a second tiled assembly, including all the devices in the second tiled assembly. Activities in different tiled assemblies resonate with one another and generate new possibilities. The resonating patterns are based on Critical Point principles that operate “the same” at every scale and on every level. When multiple possible pulse patterns and multiple possible block diagrams are co-existing together and changing continually into one another, then you’ve got
    Shimmering – the Critical condition that precedes selection of and relaxation into one particular pattern or diagram. Quad Nets support this kind of activity that is different from that supported by devices that transform inputs into outputs.

    bob kovsky (rlk “at” sonic.net)

    Like

  14. Bob said: Different kinds of stuff are inputs and the list of possible ingredients is open-ended. The controls are more important. Learn enough about the controls to get a production line going with any convenient ingredients and then you can work out the variations. Inputs are non-specific to the question of function.

    This should all be quantifiable. The controls (with parameters), the inputs, and the output (even if it is just a periodic real function). If it is open-ended, give a case with a specific set of inputs. If I want to tell you how to make Krispy Kreme glazed donuts, I can specify an algorithm for making them with a specific list of ingredients, etc..

    Even in ANN models, you can have any number of possible inputs, but this doesn’t stop people from giving helpful and simple examples with, say four inputs to a single neuron, and describing how the output relates to the inputs.

    Like

  15. Pingback: neurodudes » Blog Archive » More on “Quad Nets” (new brain/mind theory)

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s