Intro  Learning  Patterns  Real Life  Genetics  Classifiers  Biology  Neural Nets  Connectionism  Life  AI

Essays on Complex Systems Part 6 - Biology

Biology Information Relevant to Artificial Intelligence

[This is BIOLOGY.DOC, Section of Neural.doc, subsection of Complex.doc]

The Brain

Transfer information on brains here from other files.

Transplanted brain cells cause the transfer of memory

Brain cells from a quail were transplanted into a chicken, which resulted in the chicken making quail sounds. This is astounding from several different perspectives, and raises interesting questions.

First, it seems to imply that the knowledge of how to make those sounds was encoded in those brain cells such that they could function as a transplantable module.

If you transplanted a portion of one person's brain into another person, would the recipient suddenly have memories from the other person's experience? The experiment suggests Yes! [Does that mean you could conceivably repair stroke injury with a partial brain transplant from a donated organ?]

Second, the knowledge stored in the brain cells appears to have been stored there independently of context. It contained an absolute description of how to create those sounds which could function without the support of the rest of the quail's brain. Probably whatever similar support the chicken brain could provide was sufficient.

Apparently a quail and a chicken are physically similar enough that the chicken could produce the sounds.

What would happen if you transplanted the tissue into a dog? The idea of brain tissue as a context-independent storage medium may be supported by an old experiment where flatworms were trained in some task, then ground up and fed to other flatworms, who instantly became better able to learn the task than other flatworms. Maybe the knowledge was contained in chemicals that the original flatworms had produced, and which apparently weren't completely broken down when ingested.

Third, it supports the long accepted notion, that has nonetheless long baffled me, that the control of behaviors can exist in the physical wiring. It seems obvious that learning must produce physical changes, but I have always assumed that those changes were transitory, occurring within, but not to, the medium, the brain, whose basic structure remained unchanged, much as the data stored in a running computer doesn't really change the basic structure of the computer while it's there. The computer returns to its basic state when it's turned off. (But this isn't really true. The computer is in fact physically changed while it's holding the data; that's how the data is held.)

Still, it never made sense to me how instincts could exist without being somehow learned after birth, even though examples that could not possibly be learned, such as Sphex wasp behavior, abound.

But if knowledge such as how to sound like a quail can be stored in an "absolute" form, it makes sense that a brain in development can grow just the right connections at just the right strengths to store exactly that behavior from the start, much as you can manufacture a ROM for a computer with the information already in it.

It also implies that, like the ROM, but unlike RAM, the information stored in a brain probably persists after death up to the point where the brain actually starts deteriorating.

And what about the brains of people long dead that are currently preserved in storage? Though the cells are dead, if they and whatever determines the strength of the connections between the cells aren't destroyed, then it's possible that those people's memories still exist within those connections, available for retrieval if the mechanisms of function could be reactivated (which of course is the hard part).

If knowledge can be stored in absolute, or hard-wired, form in the brain, and if the physical structural changes produced by learning are not of a different nature than those, for example, present from birth as instincts, then it should be theoretically possible that a brain can be produced from the start containing any arbitrary kind of knowledge.

For example, if knowledge of geometry theorems proved in the future to be essential to survival, then there is no reason that after many years of evolution, people might be born with an innate knowledge of geometry or with at least an instinctual aptitude for it, much as they have an instinctual aptitude for language.

Maybe this is how the instinctual aptitude for language developed. Any number of species might have sufficiently complex brains to learn it, but maybe the effort is more trouble than it's worth. For example, the learning might take longer than the lifetime of an individual. Or maybe the brain resources required for it would crowd out more immediately essential functions. Any inheritable brain structures that from the start support language capabilities are a clear advantage.

So behaviors really can be inherited, which is hardly news to anybody, but now I accept it more readily with some explanation of how it can happen.

Now we get to the idea that originally inspired these few paragraphs, but which is now somewhat dwarfed by them. Behaviors can be inherited (either physically or culturally), and they, like any other physical traits, can be beneficial, detrimental, or irrelevant with respect to the individual's survival. Any physical (or physical/behavioral) characteristic can thrive as long as it isn't detrimental.

So although there is always a temptation to figure out the benefit of some trait, not everything that exists is necessarily useful. It may simply not hurt. Irrelevance probably isn't, however, completely neutral, since greater efficiency and less waste have value.

Misc

A brain stores about 10¹⁴ bits (12,500 1-gigabyte hard drives). Not sure what that means, though. If it means it can store that many passive data bits, then that's the storage limit. But if it means there are that many synapses, each synapse may have a different significance or meaning depending on what other synapses are simultaneously activated, which would dramatically increase the true storage capacity: the storage limit would be not the number of different synapses, but the number of different on/off patterns that 10¹⁴ synapses can produce, or 2 to the 10 to the 14^th power!

Left side = language, recognizing objects.

Mind reading! -- A thought is a physical phenomenon. With proper equipment and an interpretive map, you could observe them, and the interpretive maps are getting more precise!

60 Minutes, 12/2000: "Farwell (<- inventor) Brain Fingerprinting". Electrodes detect resonant activity (recognition) when a familiar stimulus is presented.

Scientific American Frontiers: Even when a subject can't tell the difference between an accurate and a false memory, the brain shows activity that differentiates between the two. A memory of something that actually happened activates 2 centers (including sensory); a memory of something that the subject only thinks happened activates only 1 center.

Scientific American Frontiers, 3/2001, "Bionic Body": Air Force researchers (Andrew Junker) and others have had success using brain waves to control computer cursor and machines. (Electrodes must be very sensitive. Junker's detector amplifies the detected waves 20,000 times.)

There is a specific brain locus of fear, the amygdule, "stimulated instantly"(?). In mammals, it is the only part of the brain fully developed at birth.

There is also a discrete pleasure center, stimulated by dopamine, triggered by circumstances.

Stimulation (action) of both is similar in various animals in analogous circumstances.

There is a chemical, oxytosin, released at times in the brain, that is intensely pleasurable and positively charges the other stimuli present at the moment. (PBS Inside the Animal Mind episode on animal emotions, 3/2000).

These mechanisms would seem to be the hard-wiring for fear-avoidance and pleasure-seeking. It is not difficult to imagine how they might be wired into the rest of the brain to monitor its activity and respond when certain patterns are detected.

Neural Connections

Brain's neuron connections aren't permanent. They connect and disconnect, possibly many times a second. (Research at Princeton.) What triggers connection and disconnection? This is like my ripping out connections that aren't helping.

A "stronger" connection is simply the result of more (duplicate) connections between the same 2 points, and the changes that result from learning (these stronger connections) are changes in the expression of genes in the neurons. (Eric Kandel, Charlie Rose 8/7/2001).

In early brain development, neurons do migrate from their origin in the neural tube to fixed locations, following (apparently) chemical pathways. They then establish important connections between general areas. After that, a large excess of specific local connections forms, which is pruned according to experiential usefulness. (PBS, Secret Life of the Brain, 1/2002).

Scientific American Frontiers 3/2001, "The Bionic Body"

Focus of show is methods of using natural neuronal capabilities in engineered ways to stimulate repair of damaged spinal cords, and methods of brain/computer/muscle interfacing.

Embryonic stem cells are undifferentiated, with the capacity to differentiate into any cell type. Early differentiation is triggered by very simple chemical switches. Washing with retinoic acid causes differentiation into nerve cells.

There are many different kinds of neurons.

An oligodendricyte (correction: oligodendrocyte) neuron wraps a portion of one or more axons in myelin sheaths, and is required for normal function.

Injected neuron cells migrate toward an injury site where they are needed.

Schwann cells from peripheral nervous system can self-repair. They also attract axon growth from other cells toward themselves; upon arrival, they coat them with myelin.

In mammals, nose nerve cells constantly renew their connections to the brain. (Does this imply that the connection is initiated from the periphery inward toward brain? Is this a general rule?)

(Implanted computer-controlled muscle electrodes activate muscles well, but proper coordination is difficult. This should be a good potential application of a program like Genghis (Rodney Brooks, MIT) that evolves according to what works well for the particular person.)

(Prolonged strenuous exercise is important to stimulate nerve regeneration, and can even trigger involuntary reflexive muscle activity, walking motions. Some knowledge of muscle activation pattern for walking is stored in spinal cord? Prolonged exercise also helped mice learn in general, and was as effective as an enriched environment.)

Language

From PBS Nova: "Wild Child", 4/2000 (originally 1994)

Cases of two children who didn't learn language in childhood due to isolation. Los Angeles 1970, "Genie", and France 1790 "Victor".

Both seemed to support Noam Chomsky's theory of a childhood window of opportunity for language. Both learned word meanings readily, but Genie was (or both were?) unable to form grammatical sentences normally. Neither child learned to speak well, though Genie's use of sign language appeared good, and she tended to use gestures a lot naturally. Suggests to me that the window of opportunity might not apply to basic symbol use. It was impossible to determine whether either child had been mentally retarded from birth.

Parsing language from sounds (speculation)

[Ref. Godel Escher Bach]

Are there hypercomplex auditory neurons?

They would act similarly to those for the eyes: in the ear, sounds are automatically pre-parsed by frequency such that particular cells fire in response to particular frequencies.

These outputs could be passed to hypercomplex cells that fire when a particular combination of the frequency detectors are firing simultaneously, such that each hypercomplex cell would detect and flag a particular pattern.

These outputs could go to "memory cells" that fire for a slightly prolonged period after receiving an input from a hypercomplex cell. (I.e. a memory of the frequency pattern just received.)

Hyperhypercomplex cells could receive inputs from the memory cells and the hypercomplex cells, and would detect 2-stage sequences of sounds, i.e. the sound just received and the one just preceding it. These outputs would go to yet another stage.

The result would be a network of end-stage neurons sensitive to particular sequences of sounds; each only fires when it detects the sequence it is dedicated to.

Into this network dangles a huge array of connections from other locations in the brain. A "main center" correlates the firing auditory neurons with other simultaneous sensory experience, and this ultimately provides sounds (words) with meaning (intrinsically experiential operational definitions).

There could be another sensor that fires when it detects no activity in any of the other circuits, (i.e. all the other circuits inhibit it, when they are active) which would be an instantaneous flag that the sound (or word, or phrase) is unfamiliar (or in the case of grammar, incorrect); this would explain why the recognition of unfamiliarity or incorrectness is instaneous, not the result of a long exhaustive linear search through all known words, sounds, etc., which would be absurd.

The key to the speed with which language is interpreted is that it is massively parallel: the same auditory stimulus is presented simultaneously to all the sensor circuits.

Each familiar pattern triggers activity along the pathways tuned to it, and the simultaneous activity along those various pathways is a thought, is the meaning of the sentence just heard.

Most of the sounds are nouns or verbs, which have easily instantaneously-summoned operational definitions and connotations that fill in the image; and some of the sounds (words) provide the logical relationships among the images just created.

Implications for my program Wtalk.cpp (most of which I decided on long ago, anyway), if it is to be based on a coherent theory of how the brain works:

1. Never bow to the pressures of data storage problems; sufficient storage will eventually become available. In the brain, each symbol does have its own storage, and every grammatical construct also does: people do not apply rules of grammar to speech they hear; they hear patterns, and every legal pattern is actually in storage somewhere. It is the only way that an incorrect or unfamiliar pattern can be instantaneously recognized.
    
2. Never worry about processing speed; that too will continually increase; if it doesn't, you must resort to parallel processing, not to clever overthought algorithms that lose the biological foundation I wanted for the program.
   
   A biologically-based program should be easier to modify as needed to accommodate changes to the theory, i.e. to evolve.
   
   A highly-indexed clever system with "artificial intelligence" algorithms (the route it seems John McCarthy favors, based, I think on an interview on PBS, show unknown) runs the risk of having to be scrapped if it turns out inadequate, and the basic problem is that it always will be inadequate because its only intelligence is whatever you built into the algorithm.
   
   A truly intelligent program must be able to work out, evolve, its own algorithms, and the higher (or maybe more technically correct, lower) level at which it can do that, the more intelligent it is.

A strange corollary to this, and my favorite, is that however complicated an intelligent program may be from the programming standpoint (infrastructure), a highly intelligent program should be conceptually very simple.

That is, it should have the tools to do the job (whatever job you anticipate for it), not the instructions for how to do it.

How to do various jobs becomes a database that it maintains itself. The all-in-one union query in Records.mdb is based (though imperfectly) on this idea; it is very powerful at extracting data and determining how to categorize it, but relatively general in its method.

Programmed Cell Death

The inferred theme is that in the body, as in evolution, the tendency is to overproduce units (cells, here), then kill off any not needed because they are either excess (too numerous) or not required in the end product. [Contrast with the theoretically more efficient method of only producing exactly what you need.]

Points

Cells all communicate with each other via signals/stimuli (chemical and electrical).

In multicellular organisms, each cell has an intrinsic self-destruct program that must continually be prevented from running, by prodding from other cells, via these signals, else the cell dies.

For example, in developing chicken embryos, if a limb is cut off, cells in the spinal cord die. They require continued signals from the limbs to stay alive.

In a developing organism, many more cells are made than are needed. The excess die.

1. There are roundworms that have exactly 959 cells, after the extra cells die off.
2. Our fingers aren't webbed because the webbing cells die off at a point in development.
3. Extra neurons are made. Neurons that make interconnections live. The rest die.

Cells removed from multicellular organism into isolation in culture die. (Really? Then how do they culture heart cells, etc?)

In some leukemias, cells become unable to self-destruct when they normally would, so the runaway growth results not from massive overproduction but from a refusal to die at a normal rate.

Possible Implications

Neural Nets: overproduce nodes and/or connections, and rip out ones you don't need.

This mechanism is missing in some analogs of living systems: e.g. bureauocracies, with negative consequences such as uncontrolled growth.

DNA

(this section reserved for combining a number of misc sections from complex.doc)

It is unlikely the DNA molecule itself sprang fully formed into existence, so it would make sense that there would have been early less efficient prototypes of it in early organisms that by our standards would barely qualify as life, protolife (RNA?, in viruses? But viruses need live hosts!).

Yet the earliest life with DNA appeared 4 billion years ago, so DNA had to develop in only 1 billion years, amazingly fast considering how little evolution there was in the following 2 billion. How?

Even more curious is that just as bacteria still survive today, such not-really-life organisms should, too. That is, if there were crude precursors to DNA that can code proteins, but inefficiently, those chemicals wouldn't have disappeared from existence. Where and what are they?

Remember how chemicals can attach themselves to DNA, blocking transcription, and others can remove the blockers to re-enable it. This is a key mode of operation of hormones in particular, and probably vitamins.

If something is being built, there can be a window of opportunity; if the needed blocker/unblocker isn't present at the right time, the structure will be different or defective, or can't be completed at all.

This principle can also explain how things can occur in a fixed sequence: a hormone can trigger the production of a second hormone, which could then turn off production of the first one and trigger the production of a third...

As they get turned on and off, different proteins are produced, presumably at the times they are needed to build something.

It's like a chemical machine; instead of gears turning other gears to trigger events, chemicals trigger production sequences of other chemicals. The key is that the chemicals produced at one time are active agents that reprogram the DNA to produce whatever set is needed next, and this is in addition to whatever non-DNA related purpose they may serve.

A chemical produced could even turn off the further production of itself!

In classifier system terms, the DNA strand is the bulletin board and the generator of messages; the chemicals produced are the messages.

[END OF BIOLOGY.DOC]