Understanding Biotic Neural Networks

From Mbscientific_wiki

[edit] Instinctive functionality of biotic neural networks - stimulus/response

Single cells demonstrate stimulus/response behavior. We can view cell functions as internally directed, e.g. replication, metabolism, structural, transport, intra-cell signaling, defense and immunity. Or, cell functions can be externally directed, e.g. motor/sensor, cell-cell signaling. So the stimulus would be that of external functions which will then garner a response triggered via internal functions.

Say the stimulus would be an external input (sense of smell of food chemical) triggering the sensory functions, and next internal functions (intra-cell signaling) which would in turn trigger motor functions (response), propelling the movement of the cell towards the food source. (Fig A left, smell and motor function are within the same cell). Now take the same stimulus-response behavior of a single cellular cluster and spread it to two differentiated cell clusters (Fig B), half handle stimulus (smell sensory neurons in red ), the rest handle the response (motor cells) .

Lets add some more complexity to the picture. Say our fictional creature (lets call it wikit) can smell food (SmF cells) and can small its enemy (SmE cells, which eats wikits). The sensory neurons will then pass the signal to a corresponding control cell cluster (C1E and C1F, which just relays the info to the motor cell cluster (Fig C, below). So if wikit smells food it'll fire the motor neurons in that direction, if it smells the enemy it'll fire away from it.

To make the figure simpler, lets combine the Smell cells SmF, SmE to a single cluster Sm (Fig D). Lets give the Attractor neurons SmF a firing value range of 0 to +SmMax and the Suppressor Neurons a value range of -SmMax to 0. Then we can combine the controller cell cluster (C1, Fig D). All C1 has to do is to integrate (add) the firing value ranges of Sm cluster and pass the result to Mo (motor) neurons. So, if the value is positive fire toward source of Sm with the proportional intensity and if Sm is negative fire away from the source, i.e. fire towards the food source and away from the enemy source.

Next, lets add some more complexity to the stimulus-response nature of our wikit. Lets give it taste sensors. It'll work just as the smell sensor-control-motor cluster did. In figure E, you'll see a parallel track for taste sensor cluster (Ta, in red ) and a corresponding control cell cluster, labeled C1,2, that is control cell cluster layer 1, track 2. And the smell control cluster becomes labeled as layer 1, track 1.

One sees an immediate problem. There is no integration between firing of C1,1 and C1,2. The motor neuron could go haywire. So in Fig F we'll add a second control layer C2 to integrate the signals of layer 1, in this case C1,1 and C1,2. It'll add the positive and negative signals and just as before proportionally fire the motor neurons towards and away from the source.

But we have another problem. The firing values are all of equal weight. In nature the firing values are dynamically weight adjusted. This requires a feedback that is either positive or negative. Say the wikit smells food and enemy, with food being the stronger signal. It inches towards the source; then wham the enemy pops up and takes out a chunk of wikit butt. Then the pain neurons will scream, sending a strong signal to the suppressor smell sensor-controller connection, amplifying those connections, thereby giving a weight, a multiplier to the enemy signal. So the next time the wikit smells food and enemy of equal magnitudes, the controller weighs the enemy signal proportionally, dwarfing the food signal and passing the information up the control chain to get the hell out of there pronto.

Another example would be comparing the smell and taste signals. Say the smell signal is long range, the taste short range. So the reward of the taste signal would be quick, sending a positive signal to the attractor taste sensor-control synapse (and/or taste/smell integrator synapse), amplifying it positively. So the next time the wikit tastes food nearby, with no suppressor signals around, it'll go for it like a bat out of hell.

So the feedback loops are a way of dynamic learning in nature, positive behavior in reinforced, negative behavior suppressed.

In nature feedback sensor-controls might take various forms, could be pain like the above example, could be pleasure (particularly yummy food, abundant food); it might be any number of things. So instead of cluttering our drawing with a whole permutation of feedback nodes (as important as they are), I'll just indicate them generally (Fig G below).

Already our wikit stimulus-response neural net can execute a fair bit of logic, and some learned behavior. Depending on the dynamically weight adjusted smell and taste values the wikit brain integrates them and decides on what to do. It can do AND/OR/NOT logic on an analog range of weighed attractor/suppressor signals. But, there's more, a lot more.

[edit] Learned functionality of biotic neural networks - heuristics

Now lets talk about memory, where training is required. Lets add in a sound stimulus, which requires the training of memory circuits for what would become an agreeable or disagreeable sound stimulus.

Pavlov's dogs put Associative Logic on the map. Say sound by itself is a benign stimulus, but it is consistently present when food is present (sounds like food?), then the logic cells would get trained over time and fire if sound is present, even if food isn't. In this example sounds form a pattern. And in normal life Pavlov's dogs would have learned to differentiate arbitrary background chatter from their owner calling them with something like "Spot, Food!". Neural networks naturally form memory circuits as the following drawings demonstrate:

So if we add in the sound sensor-control layer in parallel, we'll have

In figure H, you have trained Memory of Sound (MSnd) , something like "Wikit Food" . As before, I've also added an integrator control layer (layer 3) to add in the sound control signals (C1,3) with the control signals that we have already had from control layer 2 (C2). We are doing it because in nature bio-functional pathways have their own evolutionary morphological developmental paths, and from an engineering stand point nature does it right, other wise it'll screw up all of the other developmental pathways. So what we typically see is controls layered upon each other and integrated both functionally and morphologically.

Next we'll add in a sight control layer. Here, just like the sound control layer, we'll have a trained sight memory cluster (MSit) . And, as before I've added another control layer C4 to integrate all of the signals of preceding control layers.

If you think about it, the complexity of the executable logical profile of such a network would be rather remarkable. First, you have to consider that now you have 4 sensory input sources (S1..S4: smell, taste, sound memory, sight memory). With each signal ranging from Smax to Smin. And the affirmation of the training is a continual process, not a one shot deal. If you put all of that together, you would not expect a deterministic behavior. Rather, you would expect trial and error, give up and try again, hunt and peck, otherwise rather chaotic behavior to establish agreeable outcome in a consistent manner. And if we consider the behavior of animals, from insects, to dogs to people, that is exactly what we see. The bottom line is that the heuristic process is inherently chaotic, the learning process may or may not lead to a stable result set. Fortunately we can take the heuristic network design quite a bit further.

Suppose the wikit brain handles the smell, taste, sound and sight signal and gets something fantastically rewarding (or terribly awful). Wouldn't it be nice if it could remember the event. Well, why can't we give it an action memory complex that'll record the event details. All of the values are generated by the nodes. All we have to do is to feed the node signals into a memory neural net and the feed back signal would be the affirmation (training) signal. That is, the signals from the action nodes would normally be noise that the action memory complex would ignore, and when a strong affirmation signal is sent representing something wonderful or awful, then the complex would record it. Well, if nature can do it, we can do it to our wikit (fig J, below)

Now, we have the memory of all of these sequences of memorable actions. What if under other similar circumstances the wikit goes through slightly different action sequences and gets another affirmation (Deductive Logic). The difference between the first and the second sets of actions is the logic map of the corresponding action sets.

After all behavioral logic Kernel can be expressed as IF CONDITION THEN ACTION 1 ELSE ACTION 2.

Why can't we record that in a logic memory complex (figure K below):

So now memorized actions and their resultant outcomes can be compared. That is the logic of the actions could be affirmed across the spectra of actions. There is an internal memory of logic. We certainly have that. We can remember why and how we did what we did with respect to a given circumstance that got affirmed (turned out great, or awful) and use that as a benchmark to try it with respect to other similar circumstances. And if it didn't work out we try a different variation, and if that gets affirmed then the action and the logic gets committed to memory.

As you will see next, we aren't the only animals capable of this.

[edit] Actual case studies from the animal kingdom to drive the point home

Up until now, all of our discussions were abstract. Now lets cover some observations from the animal cognitive studies from PBS, Nova - Ape Genius:

1) In this experiment a peanut is placed in the bottom of a long glass tube that is tied up in an upright position. Never having seen this puzzle before, a chimp is to try to figure out how to get the food out. According to the Researcher Josep Call "an hour goes by and boom, they solve the problem". In the video you see the chimp going back and forth with mouth full of water, filling the tube until the peanut floats up within reach.

Lets analyze what happened. The chimp must have memory of peanut tasting good (attractor, affirmed object memory). Perhaps he has the memory of things (nuts!?!) floating on water (action memory). Then its takes him a while to put the object and action memory together, and that would provide a potential solution. (JOSEP CALL -Max Planck Institute for Evolutionary Anthropology). So you have the abstraction of object (peanut) and action (floating on water) formed in the chimp brain BEFORE he attempts the solution. And once the solution works, the chimp affirms and commits the object/action/circumstance to memory (pictorially depicted, right).

2) Lets look at another chimp behavior from the same program. It is about chimps using twigs to fish for termites. Lets break down that action into its detailed constituents: a) termite tastes good (attractor, affirmed object-termite memory), b) chimp sees termites going into the mound through holes (termite-action-1 memory), c) chimp sees termites clinging to things (termite-action-2 memory), d) chimp sees twigs stuck in holes and crevices (object-twig memory). Then some genius chimp puts these abstractions together, grabs a twig with the right size for the hole (d), puts it in the hole so termites cling to it (c, b), pulls it out full of termites, and yum he has an affirmed learned behavior (compound object-action abstraction):

So here you have a set of object-action abstractions coalesced together in a hierarchy of complexity to form an affirmed object-action abstraction that itself will be committed to memory. (http://www.janegoodall.org/chimp_central/default.asp - Jane Goodall Gombe Tanzania study site)

3) Lets add on to the order of complexity. In the same TV program you see chimps dig up a termite mound with a big stick (shovel like) then try (1a..d) to fish out the termites with twigs. Lets analyze that.

Suppose chimps try (1a..d) at first and get scant results (termites are in too deep for a twig). But through play they know big sticks (object) can dig up dirt (action). So now they add this object-action sequence prior to (1). First they dig up the mound with the big stick, exposing shallow holes. Then they try (1) to get results. If we look at it diagrammatically we get:

But as you see this requires an if-then logic. There is a sequential order to the compound action, first the mound needs to be dug up, exposing shallow holes that can be fished with twigs. So there must be some sort of logic memory present, allowing the chimp to string together action memories depending on the set of circumstances, an if-ten-else logic memory set. A note of caution: when I (we) think of logic, I project my own analytical introspection, that sort of projection must be avoided.

So, in subsequent scenarios they can try different variations of the actions to get results. As you see the layers of complexity can be compounded. (http://www.wcs-congo.org/04science/02goualougo/index.html - Goualougo - Congo chimp study site)

4) To demonstrate how new behavior can be derived from previously compounded action-logic, here's another chimp trait similar to 2, 3. a) Chimp loves bush baby meet, b) chimp sees bush baby lodge in tree holes, c) in play chimp sees sticks skewer things, d) chimp fashions sticks to skewer bush babies. Actually there is more to it, they pick out the right sticks of length and girth, strip off the leaves, nibble on one end to sharpen it. For all intents and purposes they make spears. (Fongoli, Senegal, anthropologist Jill Pruetz(Iowa State University) and psychologist Andrew Whiten (University of St. Andrews).

So we can see that abstraction complexity coalesces in a hierarchical order. Yet, from what we can tell, in nature the depth of that hierarchy is relatively shallow.

In the course of studying animal cognitive strengths researchers have used a variety of techniques. One that has yielded surprising results has been use of symbols as training tools.

First, I'll give you an example out of the same PBS Nova TV Documentary. Sally Boysen of Ohio State University asked chimps to choose between two dishes of M&Ms. One had fewer M&Ms than the other dish. The subject Sheba was to reach for the dish that was to go to another chimp Sarah, the left over dish was Sheba's. Sheba always went for the dish with more M&Ms, giving it to Sara. The impulse never allowed her to get the logic of the experiment: touch the dish with less M&Ms, giving that to Sarah and keeping the dish with more M&Ms.

The same experiment was done with the M&M dishes covered with numbers. The chimp that understood numbers got the logic of the experiment right. The dish with the smaller number goes to Sarah, the left over dish with the bigger number is the prize.

Dealing with symbols: numbers, words, pictures, etc. uses the cerebral circuits of the brain. When those control circuits are established, the complexity hierarchy can expand substantially.

Kanzi the Bonobo at the great Apr Trust is reported to be using some 350 keyboard symbols and can understand complex instructions with a vocabulary of thousands of spoken words.

( Note: You can watch above examples, plus a whole lot more on the internet, PBS, Nova - Ape Genius (hour long movie): http://www.pbs.org/wgbh/nova/apegenius/program.html )

Alex the Talking Grey African Parrot (RIP) could similarly follow complex question/answer series- See it at Youtube: http://www.youtube.com/watch?v=XcLLk-r1aSs&feature=related. (It turns out this was (is) one famous parrot.)

Similarly Dolphins can follow complex instructions(on Youtube: http://www.youtube.com/watch?v=ZwJaUFHs-C4 )

So there is evidence that animals can receive symbolic instructions and act on them at high levels of complexity. Still, the hierarchy of complexity that they can demonstrably generate is relatively shallow. This ability to think symbolically, communicate symbolically, is what sets our complex hierarchy of abstraction in motion. And that is what we shall cover in the next chapter, Evolution of the Mind. But before that I want to make one very important point.

Memories that are occasionally affirmed form short term memory. Those are manifested as internal connection strengths of the neurons in the memory circuits. However, if an action is continually affirmed, it ends up as permanently imprinting on the memory circuit, by the way of switching on a "Long Term Memory Gene" in the neurons of the memory circuit, thereby establishing long term memory. We need this to establish the next point.

Take the example of driving a stick-shift car. Many may not remember, but before automatic transmissions you had to shift gears with a stick shift. Think about the logical controls that go into that operation. (1) Your right foot has to let off the gas petal. (2) Next, your left foot has to press down the clutch. (3) Then your right hand has to shift the stick from one correct gear to the next. (4) Then your left foot is to lift off the clutch, and (5) your right foot has to step back on the gas. While all of that is going on (6) your left hand has to do the driving, corresponding to the (7) sight information that is to (8) lead you to your destination and hopefully avoid crashing the car in the process. 9) And you have to continually hone into sound information coming from the car, telling you whether to shift gears or not. So all of these controlled actions are distinct yet associated, some in succession, others simultaneous. Further, they are hierarchical in the context of getting you to your destination, i.e. the first decision is "I want to go from here to there", then "I'm going to drive along this route" and then all of the above decision/action kicks in.

One last point on this subject, heuristic behavior, when repeated and affirmed ad infinitum can appear as instinctive as pre-wired behavior. Why? Because of the switching on of the long term memory gene in the memory neurons. In fact, as daunting as driving a stick-shift car may be for a while, after years of doing the controlled actions, it becomes instinctive. One does not think about it, is not even conscious of it, he just does it. In this way, memory can turn in to instinct. Memory that is affirmed to the point of becoming literally instinctive (genetically driven) can itself serve as the affirming factor in decisions of a control circuit, i.e. behaving as and sometimes overriding externally affirmed memories.

[edit] Key

Chapter Key:

Morphological Flows, entities going through functional constructs thereby creating more complex entities with more complex functionalities:

Cell Stimulus-Response == cell differentiation ==> sensory and motor neurons

sensory neurons +motor neurons +controller neurons == life actions ==> instinctive behavior

Sensory neurons + neural memory circuits +affirming neurons == training ==> learning

Sensory neurons + memory circuits + integrator/arbitrators + controllers + motor cells == life actions ==> heuristic behavior

Inherent Reality (abstract entities and phenomenon) + heuristic networks ==imprinting ==> Perceived Reality (abstractions and an internal notion of what is physical)

[edit] links

http://www.youtube.com/watch?v=2Ei6wFJ9kCc - Hour talk on Neurocomputational models for understanding the brain circuits for learning - From GoolgleTechTalks, Dr. Mark Gluck, Rutgers U. (at YouTube)

http://www.pbs.org/wgbh/nova/apegenius/program.html - Ape cognitive and problem solving skills - hour program from PBS NOVA

- New Caledonian Crow Intelligence ( research website: http://users.ox.ac.uk/~kgroup/tools/introduction.shtml)

- Cetacean Intelligence on wikipedia: http://en.wikipedia.org/wiki/Cetacean_intelligence

- Elephant Intelligence (http://en.wikipedia.org/wiki/Elephant_intelligence)

- Dog Intelligence (http://en.wikipedia.org/wiki/Dog_intelligence).

Next: Creation of Us I

Previous: Comparative Anatomy of Biological Central Nervous Systems

Home