Archive for the ‘learning’ Category

030909 – Dennett’s competing drafts

Tuesday, September 9th, 2003

030909

Well, I think I finally begin to understand Dennett’s idea of multiple competing drafts.  What he’s getting at is very much along the lines of my flow of patterns concept.

What characterizes processes in the brain?  Lateral inhibition seems to be a fundamental process that has been adopted in the course of evolution because it has the effect of sharpening boundaries.  Hebbian learning seems also to be a fundamental process that has been adopted in the course of evolution because it has the effect of collecting similar patterns of activation together.  Taking the simple Hebbian learning paradigm as a starting point, evolution has selected a number of variants for preservation and refinement: populations of neurons vary in terms of their “plasticity per unit time” and their plasticity as a function of neurochemical modulators.

On the outputs (efferents) side, it may be that lateral inhibition is what helps resolve race conditions.  There is clearly some sort of winner take all process on the efferents side, although its scope is clearly not global because we can in fact walk and chew gum at the same time.

Suppose each neuron in the brain is connected to about 10,000 other neurons, and suppose arbitrarily that on the order in half of those connections are afferent and the other half are efferent.  Then if there are about 20 billion neurons in the brain and each receives input from 5000 other neurons, there must be about 100 trillion synapses in the brain[1] and who knows how to factor in the 200 billion glial cells that cluster around certain synapses.  This calculation makes me wonder about the distribution of glial cells.  There clearly are many fewer glial cells and there are synapses.  Something I read makes me think that the glial cells are associated with axonal synapses, but even that, at least if my estimation of 5000 axonal synapses per neuron is correct, still leaves many fewer glial cells than synapses.  About the only additional assumption I might make would be that the glia are associated with axonal synapses on cell bodies.  That might make the numbers come out right, but I don’t think so.  So I guess I’m still left puzzling over the distribution of glial cells.

Nonetheless, 100 trillion synapses is a lot of synapses.  Now go back and think about the so-called Chinese room puzzle.  The hapless person in this room is busily simulating with pencil and paper the activity of 100 trillion synapses.  It will take an awfully long time to simulate even a few seconds of brain activity.  Suppose the simulation interval (granularity) is one millisecond.  To simulate a second will require evaluating 100,000 trillion synapses.  Suppose the person is very fast and can update the state of a synapse in a second.  A year is about 30 million seconds. 100,000 trillion seconds is roughly 3 billion years.

=============== Notes =================

[1] Jeff Hawkins 2004 (p.210) estimates 32 trillion, but he doesn’t say how.  Hawkins, Jeff with Blakeslee, Sandra.  2004.  On Intelligence.  New York: Times Books, Henry Holt and Company.

030828 – Are human beings rational?

Thursday, August 28th, 2003

030828 – Are human beings rational?

My wife asked an interesting question: Do I think that human beings are inherently rational.  I think the answer is emphatically no.  Human beings have the ability to learn procedures.  One of the procedures that human beings have discovered, found useful, and passed along culturally is the procedure of logical analysis or logical thinking.  The fact that in many cases logic enables us to find good solutions to certain classes of significant problems ensures that logical analysis will be one of the procedures activated as a candidate for execution in a broad range of external circumstances and internal states.

What strikes me is that the end result of evolution selecting organisms with greater and greater ability to learn and apply procedural patterns has resulted in an organism that is capable of learning to simulate serial computations, at least on a limited scale.  Certainly it was Dennett who put this idea into my mind, but I do not believe that he arrived at this conclusion by the same path that I did.

This raises an interesting question: what kind of pattern and procedural learning capabilities are required in order to be able to simulate serial computations or, more precisely, to be able to learn and execute a logical thinking pattern?  Human beings certainly aren’t much in the way of serial computers.  We’re not fast.  We’re not computationally adept.  We don’t have a lot of dynamic memory.  Our push down stack for recursion seems to be limited to one level.  (The fact that we must use the logical thinking pattern to analyze pathological sentences like, “The pearl the squirrel the girl hit bit split,” rather than the (unconscious) language understanding pattern simply underlines this limitation on our capability for recursion.)

So, is human language ability the result of the evolution of ever more sophisticated procedural pattern learning capabilities?  Is the driving force behind the evolution of such enhanced procedural pattern learning the advantage obtained by the organisms who best understand their conspecifics?  Is this evolution’s de facto recognition that brawn being equal, better brains confer a reproductive advantage?  Now if better understanding of one’s conspecifics is the goal, language ability may just fall out automatically, because if one has a mechanism that can build a model of others, it makes it a lot easier to figure out what the other intends or is responding to.

Clearly, since the ability to take the viewpoint of another person does not manifest itself in children until some time after they have acquired at least the rudiments of language, the manifestation of the ability to take the viewpoint of another person is not a requirement for the acquisition of at least the rudiments of the language.  There seems to be a subtle distinction to be made here: when daddy says “hudie” (the Chinese equivalent of “butterfly”) and looks at, or taps, or points to a butterfly or a representation of a butterfly, something has to help the child attend to both the butterfly instance and the sound.  That something may be the emerging model of the other.  Or maybe it’s the other way around as I suggested earlier: the trick is for the parent to take advantage of his or her own model of the child in order to intuitively construct or take advantage of the situation in which both the butterfly and the sound of the word will be salient to the child.

Still, I keep coming back to the idea that the internal model of the other is somehow crucial and the even more crucial is the idea that the internal model of the other contains the other’s model of others.  As I think about it though, it seems to me that creating an internal pattern, that is to say learning a pattern, based on experience and observation of the behavior of another organism is not a capability that is uniquely human.  It would seem to be a valuable ability to have.  What seems to be special about the patterns we humans develop of other people is that we attribute to the other a self.  An or to animal can get a long way without attributing a self (whatever that means) to other creatures with which it interacts.

030826 – Parallel processing in the mind

Tuesday, August 26th, 2003

030826 – Parallel processing in the mind

I don’t know if it originated with Grossberg, but I like the concept of complementary processing streams.  Actually, he talks about it as if it always involves a dichotomy.  Could it not also be multiple (any number) parallel streams?  Certainly, the convergence of inputs from a large number of brain areas on the amygdala indicates that it’s not just dichotomous streams.

Grossberg writes as if he is describing exactly what happens—especially with his neural circuit diagrams, but the more I read, the more they seem fanciful.  Certainly there’s something missing when the diagrams only show neurons in layers 2/3, 4, and 6.

In also seems that there’s something missing from the analysis of the visual “what” pathway.  Edge and Surface processing seem very closely tied throughout.  In visual area V1, the “blob” neurons are surrounded by “interblob” neurons and in visual area V2, the “thin stripe” the neurons alternate with the “interstripe” neurons.  Surely there is some crosstalk between (among) the channels.

Grossberg uses the term “catastrophic forgetting”.  He also talks about complementary channels of processing in the brain.  And he, among others, and talks about a “where” channel to the parietal lobe and a “what” channel to the temporal lobe.  Things then get a little confused.  Part of the point of “catastrophic forgetting” is, in effect that certain memories need to get overwritten, e.g., memories of where a particular movable object is located.  In contrast other memories should not be easily forgotten.

It is not clear that the categories “easily over writable” and “not easily over writable” (or should it be “things that change often” and “things that don’t often change”?) are the same as “where” and “what”.  It’s certainly possible from an evolutionary standpoint that what and where are sufficiently essential aspects of the environment that they should be per se ensconced in genetically determined neural structures.  Nonetheless, is reasonable to ask whether what evolution has provided is also being used in ways unrelated to its evolutionarily determined functionality.

Or alternatively, given that evolution has cobbled together mechanisms capable of recording information with differing degrees of environmental permanence, it seems reasonable to suppose that the same mechanism could show up in different places; although, I am well aware that the essentially opportunistic functioning of evolution leaves open the possibility that the same function is performed in many different ways.  Still, in our environment and the environment of our animal ancestors some things change rapidly and some things don’t.

030820 – The problem of brain design

Wednesday, August 20th, 2003

030820 – The problem of brain design

David Perkins, Professor of Education at the Harvard Graduate School of Education, observes (reported by Beth Potier in Harvard Gazette) ‘It’s far easier for a group of people to pool physical effort than it is for them to effectively combine their mental energy. He illustrates this point with what he calls the “lawn mower paradox”: 10 people with lawn mowers can handily mow a lawn much faster than one, yet it’s far more difficult for the same 10 people to design a lawn mower.

‘”Many physical tasks divide up into chunks very nicely,” he says, but not so with intellectual duties. “It’s pretty hard to say, ‘Let’s make a decision together: you take part A of the decision, I take part B of the decision.'”’

So what the brain has is a large number of interconnected pattern recognition systems.  The individual systems fall into a smaller number of categories, e.g., cortical systems, cerebellar systems, etc.  System categories differ among themselves at the very least in terms of plasticity and responsiveness to various neurotransmitters and neural activity modulators.

These systems each work along the lines proposed by Natschlaeger, Markram and Maass.  This is not to say that I totally buy into their liquid state machine model, but that I do believe that the systems act as an analog fading memory (with digital overtones) and that their structure serves to project their inputs non linearly into a high dimensional space.  Different systems have different memory decay time constants, ranging from short (early visual processing, for example), to medium (audio processing for speech recognition), to long (maintaining context while reading).

I hypothesize that (at least some of) these systems become tuned (and in that sense optimized) over time to their inputs (thus improving the separation of components projected into high dimensional space) by a process approximating Hebbian learning.  This could account for the acquisition of the ability to distinguish among phonemes when learning a language.  In effect, Hebbian learning creates “grooves” into which particular stimuli are likely to fall, thus enhancing the separation of minimally differing phonemes.

030819 – Emotion and incentive

Tuesday, August 19th, 2003

030819 – Emotion and incentive

I really don’t like Joseph LeDoux’s (2002) use of the words emotion and incentive.  He uses emotion to mean just about anything that can affect synaptic plasticity, that is, he defines the term backwards.  That doesn’t work because we don’t know what can affect synaptic plasticity, but we do have a good idea of what we think emotion means.

Similarly, incentive.  To my mind an incentive is a conditional promise of reward in the future.  It takes the form, “if you do this you’ll get that.”  The term is a bit confusing in ordinary speech.  Management announces an incentive program whereby workers who overfill their quotas will receive a significant bonus.  The announcement serves as the incentive for employees to work harder.
Hans-Lukas Teuber (Chair of the M.I.T. Psychology Department while I was getting my Ph.D there) liked to tell the  story of the monkey and the “consolation prize.”  The monkey works to get a piece of banana, but when the monkey gets the piece of banana, he doesn’t eat it, he just sticks it in his mouth and holds it in his cheek.  When the monkey makes a mistake and doesn’t get a piece of banana, he eats some of the banana he was holding in his cheek.  So the monkey “rewards” himself for making a mistake.  Teuber called it a “consolation prize.”

So the (implicit) promise of “a piece of banana if you do this correctly” is the incentive (I actually would have said motivation here—LeDoux can’t because he uses motivation to mean something else).  What’s the banana then?  A reward?  Maybe, but in the context of the situation, the banana is the confirmation that the incentive was correctly understood, and that, in itself is (arguably) rewarding.

It should be rewarding, in any case, by the following argument.  It is clearly adaptive for an organism to be able to reliably predict the way the future will unfold, particularly with respect to possible events that have (can have, may have) some kind of significance to the organism.  It is even more important for an organism to be able to reliably predict the effects of a possible action

“I’ll bet that if I figure out what to do here, I’ll get a piece of banana.  Hmmm.  This looks right.  I’ll do it.  Banana!  Yes!  I was right!”

Or

“I’ll bet that if I figure out what to do here, I’ll get a piece of banana.  Hmmm.  This looks right.  I’ll do it.  No banana?  Bummer!  I didn’t get it right.  I’m gonna eat a piece of banana.”

So, back to the question: what is the banana?  In evolutionary terms, at one level, the banana is nourishment and valuable as such; in this context, however, the banana is real-world confirmation of correct understanding of (at least one aspect) of the real world.

But notice the subtlety here.  Setting aside our knowledge that correlation is not causality (which we seem to do a lot), the banana confirms the existence of a pattern: In the context of this recognizable situation it is to be expected that a problem will be presented and if I correctly figure out what the situation requires and do it, I will get some banana and if I don’t figure out what the situation requires, I won’t get any banana.

If no banana is forthcoming, what is the correct conclusion in this situation?  There are several: 1) I got it wrong (everything else is unchanged); 2) I got it right, but there are no more bananas at the moment (everything else is unchanged); 3) The pattern is incorrect: there are no bananas to be had here.  This is clearly not an exhaustive list of all the alternatives, but it does indicate that the conclusion to be drawn in the situation is by no means obvious.  This is borne out by the well-known fact that behavior patterns established by a random reinforcement pattern are more resistant to extinguishment than patterns established by a 100 percent reliable reinforcement pattern.

Again let’s look from an evolutionary standpoint: Which is more important?  Obtaining a piece of banana or knowing how to obtain a piece of banana?  If I give a man a fish, I have fed him for a day; if I teach a man to fish, I have fed him for life.

An important question for an organism is: Where is food?  The obvious next question is: How do I get there? Once these questions are answered, the next question is: Once I get there, how do I get it?  I have a feeling that in the brain these questions, or rather the answers to these questions, are intimately related.  Ultimately, an organism needs a procedural answer: What steps need to be taken in order to arrive at the desired goal?  The organism needs a sequential plan.  It makes me wonder if the parietal lobe in addition to its involvement with the representation of physical space also is involved with the representation of conceptual space.

Maybe not.  Physical space has obvious nearness relationships that conceptual space does not necessarily have.  On the other hand, George Lakoff’s arguments about the way in which meanings are derived from physical relationships may suggest that parietal lobe involvement (or, more precisely, involvement of whenever part of the brain is responsible for keeping track of the physical organization of the universe with respect to the organism) in the organization of concepts is in fact plausible.

Correlation is not causality, but from an evolutionary standpoint an organism cannot in general afford to do the necessary research to establish reliable causality.  Interestingly, human beings have acquired the ability to reason systematically and have managed in some cases to determine causality.  What is more significant, and many have remarked upon this, is that humans can transmit patterns verbally to other humans.  Not only that, patterns thus transmitted can be used by the receiver almost as if they had been directly precedent or intuited or whatever the appropriate word is to describe the way we acquire patterns.  I say “almost” because I think there must be some difference between patterns established by word-of-mouth and patterns established by other means.

I don’t think, however, that the difference is as simple as the difference between declarative and non declarative memory.  And by the way am not real happy with the use of the word declarative.  And I guess part of the reason for that is that I think some if not much of things that enter “declaratively” ends up stored “non declaratively”.  Which, I suppose, is simply to say that we don’t always consciously examine the implications of things that we hear, but those implications may end up being stored.  Perhaps this is just a matter of “stimulus generalization”, but whatever it is, it feels like a hard and fast distinction between declarative and non declarative memory is ultimately misguided.  And, in fact, studies of “priming” and in individuals whose declarative memory system is damaged in some way seem to me to imply that non declarative priming (whatever that means) also occurs in those whose declarative memory system is intact.

I suppose the argument is simply that there are two kinds of memory, but things start to feel a little too glib when people start to discuss the pathways by which information enters one memory system or the other as if in the intact organism there is no (and can be no) “crosstalk” between the two.  Maybe it’s just that in the course of reviewing the literature of the past thirty years I have concluded that where there are dichotomies it is important, even essential, not to accept them literally, for fear of missing overlooked clues to the functioning of the system.

030730 – Synaptic plasticity (Hebbian learning)

Wednesday, July 30th, 2003

030730 – Synaptic plasticity in the brain

On average, whatever that might mean, each neuron in the brain is connected to about 10,000 other neurons via synapses.  “The human nervous system is able to perform complex visual tasks in time intervals as short as 150 milliseconds.  The pathway from the retina to higher areas in neocortex alone which a visual stimuli [sic] is processed consists of about 10 ‘processing stages’.  (Thomas Nachtschlager, December 1998.  “Networks of Spiking Neurons: A New Generation of Neural Network Models” http://www.cis.tugraz.at/igi/tnatschl/online/3rd_gen_eng/3rd_gen_eng.html).  “Experimental results indicate that some biological neural systems … use the exact timing of individual spikes,” that is, “the firing rate alone does not carry all the relevant information.”  That certainly makes sense in a system that has to deal with race conditions.

“Each cubic millimeter of cortical tissue contains about 105 neurons. This impressive number suggests that a description of neuronal dynamics in terms of an averaged population activity is more appropriate than a description on the single-neuron level. Furthermore, the cerebral cortex is huge. More precisely, the unfolded cerebral cortex of humans covers a surface of 2200-2400 cm2, but its thickness amounts on average to only 2.5-3.0 mm2. If we do not look too closely, the cerebral cortex can hence be treated as a continuous two-dimensional sheet of neurons.” (Wolfram Gerstner & Werner M. Kistler.  Spiking Neuron Models: Single Neurons, Populations, Plasticity. Cambridge University Press, 2002. Chap. 9.)

“Correlation-based learning is, as a whole, often called Hebbian learning. The Hebb rule (10.2) is a special case of a local learning rule because it only depends on pre- and postsynaptic firing rates and the present state wij of the synapse, i.e., information that is easily `available’ at the location of the synapse.

“Recent experiments have shown that the relative timing of pre- and postsynaptic spikes critically determines the amplitude and even the direction of changes of the synaptic efficacy. In order to account for these effects, learning rules on the level of individual spikes are formulated with a learning window that consists of two parts: If the presynaptic spike arrives before a postsynaptic output spike, the synaptic change is positive. If the timing is the other way round, the synaptic change is negative (Zhang et al., 1998; Markram et al., 1997; Bi and Poo, 1998,1999; Debanne et al., 1998). For some synapses, the learning window is reversed (Bell et al., 1997b), for others it contains only a single component (Egger et al., 1999).

“Hebbian learning is considered to be a major principle of neuronal organization during development. The first modeling studies of cortical organization development (Willshaw and von der Malsburg, 1976; Swindale, 1982) have incited a long line of research, e.g., Linsker (1986b); Obermayer et al. (1992); Linsker (1986a); Kohonen (1984); Miller et al. (1989); MacKay and Miller (1990); Linsker (1986c). Most of these models use in some way or another an unsupervised correlation-based learning rule similar to the general Hebb rule of Eq. (10.2); see Erwin et al. (1995) for a recent review.”

030708 – Computer consciousness

Tuesday, July 8th, 2003

030708 – Computer consciousness

I begin to understand the temptation to write papers that takes the form of diatribes against another academic’s position.  I just found the abstract of the paper written by someone named Maurizio Tirassa in 1994.  In the abstract he states, “I take it for granted that computational systems cannot be conscious.”

Oh dear.  I just read a 1995 response to Tirassa’s paper by someone in the department of philosophy and the department of computer science at Rensselaer Polytechnic Institute who says we must remain agnostic toward dualism.  Note to myself: stay away from this kind of argument; it will just make me crazy.

For the record: I take it for granted that computational systems can be conscious.  I do not believe in dualism.  There is no Cartesian observer.

I do like what Rick Grush has to say in his 2002 article “An introduction to the main principles of emulation: motor control, imagery, and perception”.  He posits the existence of internal models that can be disconnected from effectors and used as predictors.

Grush distinguishes between simulation and emulation.  He states that, “The difference is that emulation theory claims that mere operation of the motor centers is not enough, that to produce imagery they must be driving an emulator of the body (the musculoskeletal system and relevant sensors).”  He contrasts what he calls a “motor plan” with “motor imagery”.  “Motor imagery is a sequence of faux proprioception.  The only way to get … [motor imagery] is to run the motor plans through something that maps motor plans to proprioception and the two candidates here are a) the body (which yields real proprioception), and b) a body emulator (yielding faux proprioception).”

What’s nice about this kind of approach is that its construction is evolutionarily plausible.  That is, the internal model is used both for the production of actual behavior and for the production of predictions of behavior.  Evolution seems to like repurpose systems so long as the systems are reasonably modular.

Grush distinguishes between what he calls “modal” and “amodal” models.  “Modal” models are specific to a sensory modality (e.g., vision, audition, proprioception) and “amodal” models (although he writes as if there were only one) model the organism in the universe.  I do not much care for the terminology because I think it assumes facts not in evidence, to wit: that the principal distinguishing characteristic is the presence or absence of specificity to a sensory modality.  I also think it misleads in that it presumes (linguistically at least) to be an exhaustive categorization of model types.

That said, the most interesting thing in Grush for me is the observation that the same internal model can be used both to guide actual behavior and to provide imagery for “off-line” planning of behavior.  I had been thinking about the “on-line” and “off-line” uses of the language generation system.  When the system is “on-line”, physical speech is produced.  When the system is “off-line”, its outputs can be used to “talk to oneself” or to write.  Either way, it’s the same system.  It doesn’t make any sense for there to be more than one.

When a predator is crouched, waiting to spring as soon as the prey it has spotted comes into range, arguably it has determined how close the prey has to come for a pounce to be effective.  The action plan is primed, it’s a question of waiting for the triggering conditions (cognitively established by some internal mental model) to be satisfied.

It is at least plausible to suggest that if evolution developed modeling and used it to advantage in some circumstances; modeling will be used in other circumstances where it turns out to be beneficial.  I suppose this is a variant of Grush’s Kalman filters argument which says that Kalman filters turn out to be a good solution to a problem that organisms have and it would not be surprising to discover that evolution has hit upon a variant of Kalman filters to assist in dealing with that problem.

It’s clear (I hope, and if not, I’ll make an argument as to why) that a mobile organism gains something by having some kind of model (however rudimentary) of its external environment.  In “higher” organisms, that model extends beyond the range of that which is immediately accessible to its senses.  It’s handy to have a rough idea of what is behind one without having to look around to find out.  It’s also handy to know where one lives when one goes for a walk out of sight of one’s home.

Okay, so we need an organism-centric model of the universe, that is, one that references things outside the organism to the organism itself.  But more interestingly, does this model include a model of the organism itself?

Certain models cannot be inborn (or at least the details cannot be).  What starts to be fun is when the things modeled have a mind of their own (so to speak).  It’s not just useful to humans to be able to model animals and other humans (to varying degrees of specificity and with varying degrees of success).  It would seem to be useful to lots of animals to be able to model animals and other conspecifics.

What is the intersection of “modeling” with “learning” and “meaning”?  How does “learning” (a sort of mental sum of experience) interact with ongoing sensations?  “Learning” takes place with respect to sensible (that is capable of being sensed) events involving the organism, including things going on inside the organism that are sensible.  Without letting the concept get out of hand, I have said in other contexts that humans are voracious pattern-extractors.  “Pattern” in this context means a model of how things work.  That is, once a pattern is “identified” (established, learned), it tends to assert its conclusions.

This is not quite correct.  I seem to be using “pattern” in several different ways.  Let’s take it apart.  The kicker in just about every analysis of “self” and “consciousness” is the internal state of the organism.  Any analysis that fails to take into account the internal state of the organism at the time a stimulus is presented is not, in general, going to do well in predicting the organism’s response.  At the same time, I am perfectly willing to assert that the organism’s response—any organism’s response—is uniquely determined by the stimulus (broadly construed) and the organism’s state (also broadly construed).  Uniquely determined.  Goodbye free will.  [For the time being, I am going to leave it to philosophers to ponder the implications of this fact.  I am sorry to say that I don’t have a lot of faith that many of them will get them right, but some will.  This is just one of many red herrings that make it difficult to think about “self” and “consciousness”.]

Anyway, when I think about the process, I think of waves of data washing over and into the sensorium (a wonderfully content-free word).  In the sensorium are lots of brain elements (I’m not restricting this to neurons because there are at least ten times as many glia listening in and adding or subtracting their two cents) that have been immersed in this stream of information since they became active.  They have “seen” a lot of things.  There have been spatio-temporal-modal patterns in the stream, and post hoc ergo propter hoc many of these patterns have been “grooved”.  So, when data in the stream exhibit characteristics approximating some portion of a “grooved” pattern, other brain elements in the groove are activated to some extent, the extent depending on all sorts of things, like the “depth” of the “groove”, the “extent” of the match, etc.

In order to think about this more easily, remember that the sensorium does not work on just a single instantaneous set of data.  It takes some time for data to travel from neural element to neural element.  Data from “right now” enter the sensorium and begin their travel “right now”, hot on the heels of data from just before “right now”, and cool on the heels of data from a bit before “right now” and so on.  Who knows how long data that are already in the sensorium “right now” have been there.  [The question is, of course, rhetorical.  All the data that ever came into the sensorium are still there to the extent that they caused alterations in the characteristics of the neural elements there.  Presumably, they are not there in their original form, and more of some are there than of others.]  The point is that the sensorium “naturally” turns sequential data streams into simultaneous data snapshots.  In effect, the sensorium deals with pictures of history.

Now back to patterns.  A pattern may thus be static (as we commonly think of a pattern), and at the same time represent a temporal sequence.  In that sense, a pattern is a model of how things have happened in the past.  Now note that in this massively parallel sensorium, there is every reason to believe that at any instant many many patterns have been or are being activated to a greater or lesser extent and the superposition (I don’t know what else to call it) of these patterns gives rise to behavior in the following way.

Some patterns are effector patterns.  They are activated (“primed” is another term used here, meaning activated somewhat, but not enough to be “triggered”) by internal homeostatic requirements.  I’m not sure I am willing to state unequivocally that I believe all patterns have an effector component, but I’m at least willing to consider it.  Maybe not.  Maybe what I think is that data flows from sensors to effectors and the patterns I am referring to shape and redirect the data (which are ultimately brain element activity) into orders that are sent to effectors.

That’s existence.  That’s life.  I don’t know what in this process gives rise to a sense of self, but I think the description is fundamentally correct.  Maybe the next iteration through the process will provide some clues.  Or the next.  Or the next.

Hunger might act in the following way.  Brain elements determine biochemically and biorhythmically that it’s time to replenish the energy resources.  So data begin to flow associated with the need to replenish the energy resources.  That primes patterns associated with prior success replenishing the energy resources.  A little at first.  Maybe enough so if you see a meal you will eat it.  Not a lot can be hard-wired (built-in) in this process.  Maybe as a baby there’s a mechanism (a built-in pattern) that causes fretting in response to these data.  But basically, what are primed are patterns the organism has learned that ended up with food being consumed.  By adulthood, these patterns extend to patterns as complex as going to the store, buying food, preparing it, and finally consuming it.

This is not to say that the chain of determinism imposes rigid behaviors.  Indeed, what is triggered deterministically is a chain of opportunism.  Speaking of which, I have to go to the store to get fixings for dinner.  Bye.

030615 – (Heterophenomnological) Consciousness

Sunday, June 15th, 2003

030615 – (Heterophenomenological) Consciousness

It’s dreary and raining and that may make people a bit depressed.  That, in turn, may make it harder for people to find a satisfactory solution to their problems.   Realizing that, I feel a bit better.  It is sometimes useful to bring something into consciousness so one can look at it.

Although we may not have access to the underlying stimulus events (constellations) that directly determine our feelings, we can learn about ourselves just as we learn about other things and other people.  We can then shine the spotlight of consciousness on our inner state and try to glean what clues we can by carefully attention.

When I say we can learn about ourselves, that is to say that we can create an internal model of ourselves and use the predictions of that model to feed back into our decision-making process.  Such feedback has the result of modifying our behavior (as a feedback system does).

The interesting thing about the internal model is that it not only models external behavior, but also models internal state.

Interesting aside: consciousness can be switched on and off.  We can be awake or asleep.  We can be “unconscious”.

What are the design criteria for human beings such that consciousness is an appropriate engineering solution?

Goals:

  • Exist in world.
  • Basic provisioning.  Homeostasis. Obtain fuel.
  • Reproduction.  Mate.  Ensure survival of offspring.

Capabilities Required to Attain Goals:

  • Locomotion.
  • Navigation.
  • Manipulation.

Functions Required to Implement Required Capabilities

  • Identification of things relevant to implementation of goals.
  • Acquisition of skills relevant to implementation of goals (note that skills may be physical or cognitive).

Capabilities Required to Support Required Functions

  • Observation.  Primary exterosensors.
  • Memory.
  • Ability to manipulate things mentally (saves energy).  This includes the ability to manipulate the self mentally.
  • Ability to reduce power consumption during times when it is diseconomic to be active (e.g., sleep at night).

Damasio (1999, p.260) says:

“Homeostatic regulation, which includes emotion, requires periods of wakefulness (for energy gathering); periods of sleep (presumably for restoration of depleted chemicals necessary for neuronal activity); attention (for proper interaction with the environment); and consciousness (so that a high level of planning or responses concerned with the individual organism can eventually take place). The body-relatedness of all these functions and the anatomical intimacy of the nuclei subserving them are quite apparent.”

Well, I have an alternative theory of the utility of sleep, but Damasio’s is certainly plausible and has been around for a while in the form of the “cleanup” hypothesis: that there is something that is generated or exhausted over a period of wakefulness that needs to be cleaned up or replenished and sleep is when that gets done.  It raises the question of whether sleep is an essential part of consciousness and self-awareness or is it a consequence of the physical characteristics of the equipment in which consciousness and self-awareness are implemented.

One talks to oneself by inhibiting (or is it failing to activate) the effectors that would turn ready-to-speak utterances into actual utterances.  In talking to oneself, ready-to-speak utterances are fed back into the speech understanding system.  This is only a slight variation of the process of careful (e.g., public speaking) speech or the process used in writing.  In writing, the speech utterance effectors are not activated and the ready-to-speak stuff is fed into the writing system.

But does it always pass through the speech understanding system?  IOW is it possible to speak without knowing what you are going to say?  Possibly.  Specific evidence: on occasion one thinks one has said one thing and has in fact said something else.  Sometimes one catches it oneself.  Sometimes somebody says you said X, don’t you mean Y and you say oh, did I say X, I meant Y.

Nonetheless, I don’t think it’s necessary to talk to oneself to be conscious.  There are times when the internal voice is silent.  OTOH language is the primary i/o system for humans.  One might argue that language enhances consciousness.  As an aside, people who are deaf probably have an internal “voice” that “talks” to them.  Does talking to yourself help you to work things out?  Does the “voice” “speak” in unexpressed signs?  When a deaf person does something dumb, does he/she sign “dumb” to him/herself?

Is there something in the way pattern matching takes place that is critical to the emergence of consciousness?  The more I think about consciousness, the less certain I am that I know what I am talking about.  I don’t think that is bad.  It means that I am recognizing facets of the concept that I had not recognized before.  That seems to be what happened to Dennett and to Damasio.  They each had to invent terminology to express differences they had discovered.

Ultimately, we need an operational definition of whatever it is that I’m talking about here.  That is the case because at the level I am trying to construct a theory, there is no such thing as consciousness.  If there were, we’d just be back in the Cartesian theatre.  Is the question: How does it happen that human beings behave as if they have a sense of self?  I’m arriving at Dennett’s heterophenomenology.  (1991, p.96) “You are not authoritative about what is happening in you, but only about what seems to be happening in you….”

To approach the question of how heterophenomenological consciousness emerges, it is essential to think “massively parallel”.  What is the calculus of the brain.  A + B = ?  A & B ?  A | B ?  A followed by B?  Thinking massive parallelism, the answer could be: All of the above.  It must be the case that serial inputs are cumulatively deserialized.  There’s an ongoing accumulation of history at successively higher levels of abstraction (well, that’s one story, or one way of putting it).  Understanding language seems to work by a process of successive refinement.  Instinctively it’s like A & B in a Venn diagram, but that feels too sharp.

The system doesn’t take “red cow” to mean the intersection of red things with cow things.  The modifier adds specificity to an otherwise unspecified (default) attribute.  So the combination of activation of “red” and the activation of “cow” in “red cow”  leads to a new constellation of activation which is itself available for further modification (generalization or restriction or whatever).  This probably goes on all the time in non-linguistic processing as well.  A pattern that is activated at one point gets modified (refined) as additional information becomes available.  Sounds like a description of the process of perception.

Massively parallel, always evolving.  It doesn’t help to start an analysis when the organism wakes up, because the wake-up state is derived from (is an evolution of) the organism’s previous life.  Learning seems to be closely tied to consciousness.  Is it the case that the “degree” of consciousness of an organism is a function of the “amount of learning” previously accumulated by the organism?

We know how to design an entity that responds to its environment.  An example is called a PC (Personal Computer).

There’s learning (accumulation of information) and there’s self-programming (modification of processing algorithms).  Are these distinguishable in “higher” biological entities?  Does learning in say mammals, necessarily involve self-programming?  Is a distinction between learning and self-programming just a conceptual convenience for dealing with Von Neuman computers?

There’s “association” and “analysis”

There is learning and there’s self programming.  Lots of things happen automatically.  Association and analysis.  Segmentation is important: chunking is a common mechanism.  Chunking is a way of parallelizing the processing of serial inputs.  Outputs of parallel processors may move along as chunks.  Given that there’s no Cartesian observer, every input is being processed for its output consequences.  And every input is being shadow processed to model its consequences and the model consequences are fed back or fed along.  Associations are also fed back or fed along.  In effect there is an ongoing assessment of what Don Norman called “affordances”, e.g., what can be done in the current context?  The model projects alternate futures.  The alternate futures coexist with the current inputs.  The alternate futures are tagged with valences.  Are these Dennett’s “multiple drafts”?  I still don’t like his terminology.  Are the alternate futures available to consciousness?  Clearly sometimes.  What does that mean?  It is certainly possible for a system to do load balancing and prioritization.  If there is additional processing power available or if processing power can be reassigned to a particular problem.  Somehow, I don’t think it works that way.  Maybe some analyses are dropped or, more likely, details are dropped as a large freight train comes roaring through.  Tracking details isn’t much of a problem because of the constant stream of new inputs coming in.  Lost details are indeed lost, but most of the time, so what?

Language output requires serialization as do certain motor skills.  The trick is to string together a series of sayings that are themselves composed of ordered (or at least coordinated) series of sayings.  Coordination is a generalization of serialization because it entails multiple parallel processors.  Certainly, serial behavior is a challenge for a parallel organism, but so are all types of coordinated behavior.  Actions can be overlaid (to a certain extent, for example: walk and chew gum; ride a horse and shoot; drive and talk; etc.) Week can program computers in a way that evolution cannot hardwire organisms.  On the other hand, evolution has made the human organism programmable (and even self programmable).  Not only that, we are programmable in languages that we learn and we are programmable in perceptual motor skills that we practice and learn.  Is there some (any) reason to think that language is not a perceptual motor skill (possibly writ large)?

Suppose we believe that learning involves modifications of synaptic behavior.  What do we make of the dozen or so neurotransmitters?  Is there a hormonal biasing system that influences which transmitters are most active?  Is that what changes mode beyond just neural activity in homeostatic systems?  Otherwise, does the nature of neuronal responses change depending on the transmitter mix, and can information about that mix be communicated across the synaptic gap?  These are really not questions that need to be answered in order to create a model of consciousness (even though they are interesting questions) but they do serve as a reminder that the system on which consciousness is based is only weakly understood and probably much more complicated even than we think (and we think it’s pretty complicated).

I seem to have an image — well, a paradigm– in mind involving constraints and feature slots, but I don’t quite see how to describe it as an algorithm.  This is a pipelined architecture, but with literally millions of pipelines that interact locally and globally.  The answer to “what’s there?” or “what’s happening?” is not a list, but a coruscating array of facets.  It is not necessary to extract “the meaning” or even “a meaning” to appreciate what is going on.  A lot of the time, nothing is “going on”; things are what they are and are not changing rapidly.

Awareness and attention seem to be part of consciousness.  One can be aware of something and not pay attention to it.  Attention seems central — the ability to select or emphasize certain input (and/or output) streams.  What is “now”?  It seems possible to recirculate the current state of things.  Or just let them pass by.  Problem: possible how?  What “lets” things pass by?  The Cartesian observer is so seductive.  We think we exist and watch our own private movie, but it cannot happen that way.  What is it that creates the impression of “me”?  Yes, it’s all stimulus-response, and but the hyphen is where all the state information is stored.  What might give the impression of “me”?  I keep thinking it has something to do with the Watslawyck et al. [(1969(?) The Pragmatics of Human Communication] idea of multiple models.  This is the way I see you. This is the way I see you seeing me.  This is the way I see you seeing me seeing you.  And then nothing.  Embedding works easily once: “The girl the squirrel bit cried.”  But “The girl the squirrel the boy saw bit cried” is pathological.

As a practical matter, if we want to create an artificial mind, we probably want to have some sort of analog to the homunculus map in order to avoid the problem of having to infer absolutely everything from experience.  That is, being able to refer stimuli to an organism centric and gravity aware coordinate system, goes a long way towards establishing a lot of basic concepts: up-down, above-below, top-bottom, towards-away, left-right, front-back.  Add an organism/world boundary and you get inside-outside.  I see that towards-away actually cheats in that it implies motion.  Not a problem because motion is change of position over time and with multiple temporal snapshots (naturally produced as responses to stimuli propagate through neural fields), motion can be pretty easily identified.  So that gets things like fast-slow, into-out of, before-after.  We can even get to “around” once the organism has a finite extent to get around.

What would we expect of an artificial mind?  We would like its heterophenomenology to be recognizably human.  What does that mean?  Consider the Turing test.  Much is made of the fact that certain programs have fooled human examiners over some period of time.  Is it then the case that the Turing test is somehow inadequate in principle?  Probably not.  At least I’m not convinced yet that it’s not adequate.  I think the problem may be that we are in the process of learning what aspects of human behavior can be (relatively) easily simulated.  People have believed that it is easy to detect machines by attempting to engage them in conversation about abstract things.  But it seems that things like learning and visualization are essential to the human mind.  Has anyone tried things like: imagine a capital a.  Now in your imagination remove the horizontal stroke and turn the resulting shape upside down.  What letter does it look like?

Learning still remains intransigent problem.  We don’t know how it takes place.  Recall is equally dicey.  We really don’t seem to know any more about learning skills that we do about learning information.  We’re not even very clear about memorizing nonsense syllables for all the thousands of psychological experiments involving them.  Is learning essential to mind?  Well, maybe not.  Henry can’t learn any conscious facts, and he clearly has a mind (no one I know of has suggested otherwise).  Okay, so there could be a steady state of learning.  The ability to learn facts of the kind Henry Molaison couldn’t learn isn’t necessary for a mind to exist.  We don’t know whether the capacity for perceptual motor learning is necessary for a mind to exist.  Does a baby have a mind?  Is this a sensical question?  If not, when does it get one?  If so when did it develop?  How?

It begins to feel like the problem it is to figure out what the question should be.  “Consciousness” seems not to be enough.  “Mind” seems ill-defined.  “Self awareness” has some appeal, though I struggle to pin down what it denotes: clearly “awareness” of one’s “self” but then what’s a “self” and what does “awareness” mean?  Surely self-awareness means 1)  there is something that is “aware” (whatever “aware” means”), 2) that thing has a “self” (whatever “self” means), and that thing can be and is “aware” of its “self”.  A person could go crazy.

Is this a linguistic question — or rather a meta linguistic question: what does “I” mean?  What is “me”?  In languages that distinguish a “first person” it would appear that these questions can be asked.  And by the way, what difference does it make if the language doesn’t have appropriate pronouns and resorts to things like “this miserable wretch begs forgiveness”?  Who’s doing the begging?  No.  That’s not the question.  What’s doing the begging. heterophenomenologically, it doesn’t matter if I say it referring to myself or referring to another person.  Except that it has for me a special meaning when it refers to “my self” and that special meeting is appreciated, that is, understood, by others hearing “me” say it.

I don’t know anything about children learning what “I” and “me” refer to.  I remember reading something about an (autistic I think) child who referred to himself in the third person, for example: “he’s thirsty”

Consciousness seems to require inputs.  That is, one cannot just “be conscious” rather one must “be conscious of” things.  That sounds a bit forced, but not if it is precisely the inputs that give rise to consciousness.  No inputs, no consciousness.  Something in the processing of inputs gives rise to the heterophenomenological feeling of being conscious.

Does self-awareness have to do with internal models?  Does the organism have an internal model of the universe in which exists?  Does that model include among the entities modeled, the organism itself?  And is it necessary that the model of the organism include a model of the internal model of the universe and its component model of the organism?  It may not be an infinite series.  In fact it can’t be.  The brain (or any physical computer) has finite capacity.

But doesn’t a model imply someone or something that makes use of the model?  We keep coming back to metaphors that encourage the Cartesian fallacy.

Let’s think computer systems design.  Hell, let’s go all the way, let’s think robot design.  The robot exists in a universe.  The robot’s program receives inputs from its exteroceptors about the state of the universe and its inputs, suitably processed, are abstracted into a set of signals representing the inputs — in fact representing the inputs over a period of time.  The same thing is happening with samples representing the interoceptors monitoring the robot’s internal mechanical state: position of limbs, orientation, inertial state (falling, turning, whatever), battery/power level, structural integrity.

On the goals side, based on the internal state, the robot has certain not action triggers, but propensity triggers.  For example: When the internal power level or the internal power reserves fall below a particular threshold, the goal of increasing power reserves is given increased priority.  But we do not assume that the robot has a program that specifies exactly what to do in this state.  The state should trigger increased salience (whatever that means) and attention to things in the current environment that are (or have been in the learned past) associated with successful replenishing of power reserves.

At all times, the important question is: “what do I do now?”  The answer to this question helps determine what needs “attention” and what doesn’t need “attention”.  As a first approximation, things not “associated” with current priority goals are not attended to.  Well, it’s not quite a simple as that.  Things that don’t need attention, even though associated with an ongoing task (like walking or driving) don’t get attention processing.  Attention is the assignment of additional processing power to something.  Additional processing power can boost the signal level to above the consciousness threshold and can reduce the decay rate of attended signals.

No one has succeeded in explaining why heterophenomenological evidence indicates that people feel “conscious” sometimes and when they don’t feel “conscious” they don’t “feel” anything and they shift back and forth.  It’s a processing thing.  If I close my eyes and lie quietly, I’m not asleep.  I still hear things.  I can still think about things.  So consciousness can be turned on and off in the normal organism.  What’s going on here?  Understanding the neural connections won’t do it.  We would need to know what the connections “do”; how they “work”.

Sleep.  In effect, the organism can “power down” into a standby state (for whatever evolutionary reason).  If the threshold for external events is set high, most of them won’t make an impact (have an effect).  It’s like a stabilized image on the retina.  It disappears — well, it fades.  No change equals no signal.  If there’s nothing to react to, the organism, well, doesn’t react.

If outside inputs are suppressed, where do daydream inputs come from?  Not a critical question, but an interesting one.  Somebody pointed out that so-called “dream paralysis” is a good thing in that it keeps us from harming ourselves or others in reaction to dream threats or situations.

030604 – Wants (more)

Wednesday, June 4th, 2003

030604 – Wants (more)

Could it be that the fundamental nature of wanting is IRMs (innate releasing mechanisms) and FAPs (fixed action patterns)?  Certainly IRMs and FAPs have a long and honorable evolutionary history.  There is certainly reason to say that lower animals are a soup of IRMs and FAPs.  Why not higher animals, too?  If I don’t know what I want until I see what I do, is that just a way of saying that I don’t have direct access to my IRMs?  Or is that just silly?

And what does it make sense for evolution to select as generic wants to be activated when there’s nothing pressing?  How about something like

–    Learn something new
–    Acquire a new skill (What’s a skill?  A complex perceptual motor pattern?)
–    Practice an acquired skill
–    Think about something interesting (What’s interesting?)
–    Stimulate yourself
–    Play with the external world (What’s play?)

You can’t have a theory of consciousness without including:

–    Wanting (approach)
–    Absence of wanting / indifference
–    Negatively directed wanting / wanting not (avoidance)
–    Learning
–    Skill acquisition (Perceptual / Motor Learning)
–    Imitation (human see, human do)
–    Pleasure / Satisfaction
–    Pain / Frustration
–    Salience / Interest
–    Metaphor

[Is this my own rediscovery of what Jerry Fodor (and presumably many others) call propositional attitudes?  Some of the items are, but others are not.]

If you stick out your tongue at a baby, from a very early age, the baby will imitate the action.  But the baby can’t see its tongue, so how does it know what to do.  It’s a visual stimulus, but the mirroring is not visual.  Now, it’s possible that a baby can see its tongue, if it sticks it out far enough, but unless the baby has spent time in front of a mirror, there’s no reason to believe the baby has ever seen its own face head-on (as it were).

Children want to do what they see their older siblings doing.  It seems to be innate.  It would seem to be rather peculiar to argue that children learn to want to imitate.  But how does a child (or anybody, for that matter) decide what it wants to imitate now?  There’s “What do I do now?”  “Imitate.” and “what do I want to imitate?”

A “high performance skill” (Schneider 1985): more than 100 hours of specialist training required; substantial numbers of trainees fail to acquire proficiency; performance of adepts is qualitatively different (whatever that means) from that of non-adepts.  There are lots of examples of high performance skills.  People spend lots of time practicing sports, learning to work machinery, etc.  Why?  Improving a skill (developing a skill and further developing it) is satisfying.  Does general knowledge count as a skill?  Can we lump book learning with horsemanship?

What about Henry Molaison, whose perceptual motor skills improved but he did not consciously recognize the testing apparatus?  Not really a problem.  There’s a sense in which the development of perceptual motor skills is precisely intended to create motor programs that don’t require problem solving on-the-fly.  Ha!  We can create our own FAPs!  [This is like blindsight.  Things that do not present themselves to the conscious-reporting system (e.g., Oh, yeah, I know how to do this pursuit rotor thing.) are available to be triggered as a consequence of consciously reportable intentions and states of mind (e.g., I’m doing this pursuit rotor thing.).  So part of what we learn to do consciously is learned and stored in non-reportable form (cf. Larry Squire’s papers on the topic).  But in the case of blindsight, some trace of detectablility is present.]

But if we can create our own FAPs, we must also create our own IRMs.  That means we have to create structures (patterns) that stretch from perceptions to behaviors.  Presumably, they are all specializations.  We create shortcuts.  If shortcuts are faster (literally) then they will happen first.  In other words, the better you get at dealing with a particular pattern, the more likely that pattern will be able to get to the effectors (or to the next stage of processing) first.   Is that what lateral inhibition does?  It gives the shortcut enough precedence to keep interference from messing things up.  In other words, lateral inhibition helps resolve race conditions.  [“Race conditions” reminds me that synchronous firing in the nervous system proceeds faster than anything else.]

Consciousness (whatever that means, still) is a tool for learning or for dealing with competing IRM/FAPs.  What do I mean “dealing with”?  Selecting among them, strengthening them or weakening them, refining them.  (There.  I got revising which was close, but not quite correct.  I typed it and then I got refining which was le mot juste (and it varies only in two consonants /f/ for /v/ which is only unvoiced for voiced and /s/ for /n/ which have no connection as far as I can tell).  [Find research on tip-of-the-tongue (TOT) phenomena.]

TOT: “partial activation” model v. “interference” model.  It seems to me that these are the same thing in my model of shortcuts and races.

The problem of observational learning: assuming that human infants are primed to learn from observation (or is it that they are primed to imitate actions they perceive, particularly humanish actions?).  Suppose moreover that humans have a way of segmenting perceptions and associating the segments.  Be real careful here: Marr suggests that visual inputs get taken apart and pieces processed hither, thither, and yon.  They never need to get put together because there’s no Cartesian observer.  So associations between percepts and imitative action patterns are spread out (multi-dimensional, if you will) without the need to segment the patterns any more than they are naturally.

As Oliphant (1998? submitted to Cognitive Behavior, p.15) says, “Perhaps it is an inability to constrain the possible space of meanings that prevents animals from using learned systems of communication, even systems that are no more complicated than existing innate signaling systems.”

Oliphant also says (1998? submitted to Cognitive Behavior, p.15), “When children learn words, they seem to simplify the task of deciding what a word denotes through knowledge of the existence of taxonomic categories (Markman, 1989), awareness of pragmatic context (Tomasello, 1995), and reading the intent of the speaker (Bloom, 1997).”  [Are some or all of these consequences of the development of attractor basins?  Is part of the developmental / maturational process the refinement of the boundaries of attractor basins?  Surely.]

It begins to feel as if imitation is key.  Is the IRM human-see and the FAP human-do?  Refinement is also the name of the game: patterns (input and output) can be refined with shortcuts.  There are innate groundings.  The innate groundings are most likely body-centric, but then again, imitation has an external stimulus: the behavior to imitate.

I’ve been finding lots of AI articles about cognitive models that use neural networks.  Granting that they are by nature schematic oversimplifications, there is one thing that seems to characterize all of them, and it’s something that has bothered me about neural networks all along: they assume grandmother-detectors.  That is, they have a set of input nodes that fire if and only if a particular stimulus occurs.  The outputs are similarly specific: each output node fires to signal a specific response.  Of course, this is pretty much a description of the IRM / FAP paradigm and, following Oliphant (1998?), the interesting problems seem to be happening in the system before and after this kind of model.

There are two easy ways of initializing a neural network simulation: set all weights to zero or set the weights to random values.  But assuming that what goes on in the brain bears at least some resemblance to what goes on in a neural network simulation, it seems clear that evolution guarantees that neither of these initialization strategies is used ontogenetically.  Setting all connection strengths to zero gives you a vegetable, and setting connection strengths randomly gives you a mess.  Surely evolution has found a better starting point.  [Cf. research on ontogenetic self-organization.]

One researcher’s baby is another researcher’s bathwater.  Hmmm.  Ain’t thinking grand?

Given that there aren’t grandmother detectors [although there are some experiments that claim Raquel Welch detectors, I think] and that there are not similarly specific effectors, we are back to Lashley’s problem of serial behavior.  What keeps the pandemonium from just thrashing?  I keep coming back to a substrate of plastic (i.e., tunable, mutable, modifiable, subsettable, short-cuttable) IRMs and FAPs.  Babies don’t get “doggie” all at once.  There seems to be a sort of bootstrap process involved.  Babies have to have enough built in to get the process started.  From that point on, it’s successive refinement.

I wrote “invisible figre” then stopped.  My intention had been to write “invisible fingers”.  I had been reading French.   I don’t know for [shure] sure how the ‘n’ got lost, but the “gre” would have been a Frenchified spelling and “figre” would not have had the nasalized consonant that would have (if pronounced in French) produced “fingres”.

All these little sensory and motor homuncuili in the cortex—maybe what they are telling us is pretty much what Lakoff was saying, namely that our conception of the universe is body-centric.  Makes good sense.  That’s where the external universe impinges upon us and that’s where we impinge on the external universe.  I couldn’t think of a better reference system.

Chalmers (The Conscious Mind, 1996) believes that zombies are logically possible because he can imagine them.  He believes that a reductionist explanation of consciousness is impossible.  It is certainly true that it is a long jump from the physics of the atom to the dynamics of Earth’s atmosphere that give rise to meteorological phenomena, but we don’t for that reason argue that a reductionist explanation is impossible.  Yes, it’s a hard problem, but it requires poking one hell of a big hole in our understanding of physics to believe that a scientific explanation is impossible and therefore consciousness must be supernatural.  I don’t think I want to read his book now.  I feel it will be like reading a religious tract arguing that evolution is impossible.  my Spanish Literature Professor Juan Marichal once observed, a propos a book written by a Mexican author who had conceived a virulent hatred for Cortez (from a vantage point 400 years after the conquest of Mexico) that it is possible to learn something even from works written by people who have peculiar axes to grind.  So maybe sometime I’ll revisit Chalmers, but not now.

Antonio Damasio (1999, The Feeling of What Happens: Body and Emotion in the Making of Consciousness.) The trouble with neural nets is often that they have no memory other than the connection weights acquired during training.  A new set of data erases or modifies the existing weights rather than taking into account what had been learned thus far.  Learning from experience means that there is some record of past experience to learn from.  Of course, that may just be the answer: memory systems server to counterbalance the tendency to oscillate or to go with the latest fad.  If a new pattern has some association with what has gone before, then what has gone before will shape the way in which the new pattern is incorporated.  If there is a long-term record of an old pattern, it will still be available at some processing stage even if the new pattern becomes dominant at some other processing stage.  So, it may not be necessary to solve in a single stage of processing the problem of new data causing forgetfulness.

Learning has to be going on at multiple levels simultaneously.  Alternatively, there are nested (layered? as in cortical layers) structures that feed information forward, so some structures learn from direct inputs and subsequent structures learn from the outputs of the structures that get direct inputs and so on.

Antonio Damasio (1999) has given me the idea that will, I think, account for wanting.  Homeostasis.  The argument goes like this.  In unicellular organisms, homeostasis doesn’t have a lot of ways to operate.  When an organism becomes mobile, homeostatic processes can trigger behaviors that with better than chance probability (from an evolutionary standpoint) result in internal state changes that serve to maintain homeostasis.  In effect, evolution favors behaviors that can be triggered to achieve homeostatic goals.

In complex organisms, there are homeostatic mechanisms that work on the internal environment directly, but there are some internal environment changes for which it is not possible to compensate adequately by modifying the internal environment directly.  Thence, hunger.  Hunger is how we experience the process that is initiated when homeostatic mechanisms detect an insufficiency of fuel.  (Actually, it’s probably more sophisticated than that—more like detection of a condition in which the reserve of fuel drops below a particular threshold—and maybe there are multiple thresholds, but the broad outline is clear.)

All organisms have phylogenetically established (built-in) boot processes for incorporating food.  In mammals, there is a rooting reflex and a suckle reflex.  Chewing (which starts out as gumming, but who’s worrying?) and swallowing are built-ins as well.  But those only help when food is presented.  Problem: how to get food to be presented?  Well, if food is presented before hunger sets in, it’s not a homeostatic problem.  If not, homeostatic mechanisms switch the organism into “need-fuel mode”.  In “need-fuel” mode, organisms do things that tend to increase the likelihood that fuel will become available.  Babies fuss, and even cry, sometimes lots and loudly.

Pain is another place where internal homeostatic processes intersect with the external universe.  Pain is how we experience the process that is initiated when homeostatic sensors detect deviations from damage to internal stability that arise from a physical process (heat, cold, puncture, etc.).  Again, evolution has sophisticated the process somewhat.  The pain process arises when a threshold condition is passed.  Pain does not wait for serious damage to take place, pain is triggered when it’s time to take action to prevent serious damage.

Pain actually has to be a bit subtle, too.  Some pain may and should be ignored.  If fight is an alternative to flight, then fight arguably ups the threshold for debilitating pain.

There are other obvious situations in which homeostatic considerations require some action with respect to the outside world.  Urination and defecation are two.  Similarly, vomiting (with its warning homeostatic signal, nausea).

Our wanting, then, has its origin as the experience of a process that responds to some (serious or prospectively serious) homeostatic imbalance.

As an aside, I want to propose that one of the characteristics that distinguishes reptiles from mammals is that when a reptile is in reasonable homeostatic equilibrium, it does nothing.  When a mammal is in the same state, it does something—explores its environment, plays, writes poetry, etc.  In the most general terms, it sets out to learn something.  This characteristic arguably confers at least a marginal advantage to animals that possess it, viz. it is possible that something learned in the absence (at the time) of any pressing need will turn out to be valuable in dealing with future situations in which there will be no opportunity to learn it.  So, the concept of homeostasis has to be broadly construed.

My central point, however, is that ultimately our wants, wishes, desires, dislikes, disgusts, and delights all refer to internal homeostatic processes.  The fact that there are so many distinguishable variants of wanting suggests to me that the many shades of our experience reflect the many kinds of homeostatic processes that have been phylogenetically established in our brains and bodies, each presumably for the most part having proved advantageous over evolutionary time.