Daniel Dennett's The Intentional Stance

The Intentional Stance

I recently finished reading Daniel Dennett's The Intentional Stance. It is more or less a collection of his early articles about the intentional stance, with some editing for cohesion.

It's a pretty slow read; just 350 pages but his prose requires careful attention. I found it worth the time. He actually explains the main idea in the first few pages. Dennett has never been one of those philosophers who talks for a dozen chapters before he gets to his real ideas.

Roughly speaking, he defines three levels of analysis:

• The physical stance. All things are made up of physical material (atoms and molecules, usually) and obey physical laws. This applies to anything.
• The design stance. Some things behave as though selected for a particular purpose. This applies to artifacts and products of evolution - things that are designed, whether by "real" designers or blindly by evolution.
• The Intentional Stance. Some things behave as if they held beliefs and goals. This stance applies to intelligent living things.

He then defends why stances other than the physical should even be needed or useful. Roughly, because they predict so well from sparse data.

Dennett gives the example that, given the phone call "Honey, could you pick up a gallon of milk on the way home?", we can predict a very high likelihood that some ten or twenty minutes later, the man on the phone will drive a car into the driveway and enter the house carrying a gallon of milk (with some assumptions by context, eg assuming they have a driveway). A "martian" who uses exclusively the physical stance might make the same prediction by number-crunching through an overwhelming deluge of physical data, laboriously calculating the non-collision of the man's vehicle with other vehicles at intersections and so forth. But obviously the martian is doing it the hard way.

On mistakes

In "Making Sense Of Ourselves" and "When Frogs (and others) Make Mistakes", he considers the intentional stance when agents make mistakes. By which he means acknowledged mental failures, like a boy who sells lemonade for $0.12 giving a customer $0.11 change for a quarter. He didn't think that 11 + 12 = 25 and he wasn't trying to steal $0.02, he just made a mistake.

Dennett finds, roughly, that mistakes are gaps in the applicability of the intentional stance, much like the design stance will fail to predict when a broken alarm clock will ring. And he find that the intentional stance remains viable nonetheless, just as the design stance remains useful even if some alarm clocks do break.

Interesting quote on a side-topic:

[Quine] has been unable to give a detailed and realistic example, not because his thesis is false but because it postulates the possibility of a case poised on a knife-edge - two [possibilities] such that all the available relevant information fails to favor one over the other. The actual world abhors knife-edges even more than vaccuums. Almost inevitably grubby details begin to mount faster on one side than the other [. . .]

Beyond Belief

In "Beyond Belief", the longest chapter, Dennett takes issue with the "propositional attitudes" model and finds that what we have beliefs about are not objects in the real world, but notional objects. The notional objects are nevertheless generally tightly bound to real objects.

He also takes issue with the de dicto / de re dichotomy. He gives the example of a prank that he contemplated as a youth but never actually committed. There were two very nearly identical restaurants, and the prank would have been to have dinner with his friends at one restaurant, slip one of his friends a mickey, and drive him to the identical other restaurant, arranging everything identically. Thus the status of all of his friend's beliefs about "the restaurant" would be sensitive to whether they were about "the restaurant I'm in" or "the restaurant I went to tonight".

Dennett also uses the example of a belief about "the shortest spy in the KGB". This of course refers to an actual person (given that there are no exact ties and at least one KGB spy), yet the actual person is practically undiscoverable. Consider the belief that this person is female. The truth value of this belief can change from moment to moment as the KGB roster secretly changes, and without any "ordinary" change in the believer. Dennett borrows Geach's term "Cambridge change" to describe this sort of change, "like that change that befalls you when you suddenly cease to have the property of being closer to the north pole than the oldest plumber born in Utah"

Building on these examples, Dennett makes a persuasive argument that the de dicto / de re dichotomy misses something crucial. ISTM de dicto and de re still remain useful shorthands, as long as one that they abbreviate a potentially more complex situation.

Intentional systems in cognitive ethology / Interpreting Monkeys, Theorists, and Genes

He reviews some experiments on vervet monkeys - civilized experiments, I want you to know, like playing fake monkey alarm calls and seeing how the other monkeys react. He proposes that we should design experiments with an eye on the intentional stance.

He also takes issue with Lewontin and Gould's stance against "adaptionism" and "The Panglossian Paradigm", finding that they have more in common with B F Skinner than they think.

Another great quote on a side-topic, this time in a footnote and borrowed from Kahnemann:

I shall always treasure the visual memory of a very angry philosopher, trying to convince an audience that `if you believe that A and you believe that if A then B, then you must believe that B'. I don't know whether he had the moral power to coerce anyone to believe that B, but a failure to comply does make it quite difficult to use the word `belief', and that is worth shouting about.

Retreading

In "Error, Evolution, and Intentionality" he largely revisits themes from the earlier chapters. Apparently other philosophers had misunderstood him, and he is clarifying himself and rebutting them.

Sometimes it takes years of debate for philosophers to discover what it is they really disagree about. Sometimes they talk past each other in a long series of books and articles, never guessing at the root disagreement that divides them. But occasionally a day comes when something happens to coax the cat out of the bag. "Aha!" one philosopher exlaims to another, "so that's why you've been disagreeing with me, misunderstanding me, resisting my conclusions, puzzling me all these years!"

There's enough new material here that I don't think I wasted my time, but it doesn't lend itself to much of a summary other than "largely revisits".

Fast Thinking

This chapter is somewhat anomalous for this book. Here Dennett opines that maybe slow AI isn't AI at all, because thinking fast enough to keep up with events as they happen is a crucial part of intelligence.

Mid-term examination

Largely his thoughts on philosophy "as she is done". This is the first time I've seen Dennett address this subject. Sometimes I have thought that he must be awfully frustrated with some of his fellow philosophers, who seem to misunderstand points that this layman easily grasps. But he tackles the subject without a cross word.

Priority squabbles [. . .] in philosophy [. . .] tend to take on the air of disputes among sailors about who gets credit for first noticing that the breeze has come up

git-svn with less pain

How to make a svn clone with git-svn without spending 10 hours

I recently had the mixed experience of cloning the Rosegarden source from its SVN repo with git-svn.

First thing: Do not believe anyone who tells you to just launch:

git svn clone -s REPO-URL 

If somebody tells you to do that, give them a mean glare from me. It will take hours and hundreds of megabytes. I quit after a total of maybe 10 or 11 hours, after maybe a dozen restarts with no end in sight. It had used about 350 megabytes of disk, and it wasn't anywhere near finished downloading.

Second thing: You're going to have to find the latest SVN revision number by hand. At least, I found no way to do it within the git workflow¹. You can find it remotely via svn but if you wanted to use svn you wouldn't be doing this.

What worked for me is:

git svn clone -r $REV:HEAD -s $URL $DIR

where $REV is a fairly recent SVN revision number; $URL and $DIR are obvious.

At the risk of making my own post redundant, this Stackflow thread showed me how to do it.

Footnotes:

¹ I tried just initting the directory with git-svn and launching various informational git-svn commands from in it; none worked.

Simulated evolution of dark matter supports PDM

Simulated evolution of dark matter supports Parallel Dark Matter

This is going to be a short one. I just read a link that tends to support PDM. The story was actually out a few weeks ago but I didn't find it until just now.

Dark matter is similar to that of visible matter. ScienceDaily. Retrieved August 17, 2011, from http://www.sciencedaily.com/releases/2011/07/110721102021.htm

Summary

A team of astrophysicists and cosmologists found that "large cosmic structures made up of dark and normal matter evolve along the same lines."

PDM predicts this, and most other dark matter theories don't.

Journal Reference

M. Demianski, A. Doroshkevich, S. Pilipenko, S. Gottloeber. Simulated evolution of the dark matter large-scale structure of the Universe. Monthly Notices of the Royal Astronomical Society, 2011; 414 (3): 1813 DOI: 10.1111/j.1365-2966.2011.18265.x

Crazy idea? Parallel Dark Matter

General background

A lot of ink and electrons have been spilled over the question of what dark matter is. A theory about it has been kicking around in my head for a few years now. It could be completely wrong. But I haven't seen it out there before, and it doesn't obviously fly in the face of observations, so I'm putting it out there. If it's wrong, so be it.
I think most of my readers are already familiar with dark matter, so I'm just linking to the wiki entry. But I do want to highlight a few points:

• Whatever dark matter is, it's probably not an undifferentiated swarm of particles (WIMPs usually). This is known from the bullet cluster (1E0657-556) and similar observations.
• There's not enough gravitational lensing for it to be mostly or all MACHOs.
• And despite some neat fits to observations, theories like MOND that modify gravity have extreme theoretical difficulties.
• While dark matter is generally co-located with visible galaxies, there are seemingly random exceptions in both directions. Eg, NGC 4736 and VIRGOHI21.

Parallel Dark Matter

Probably the easiest way to explain Parallel Dark Matter is to start by pointing at braneworld cosmology. That doesn't mean PDM is committed to Braneworld cosmology, just that they fit neatly.
Braneworld cosmology says that everything we see, everything in the visible universe, is actually stuck to the surface of a brane. The only things that aren't stuck to it are gravitons, and that's used to explain why gravity is so weak. One braneworld theory, the Ekpyrotic universe, proposes two branes which collided and more-or-less made the Big Bang.
PDM proposes that there are 6 branes including ours, all about equally full. It proposes that dark matter is the matter of other branes, whose gravity alone escapes those branes.

A few immediate objections

But branes+gravitons is more complex than that

Q: Gravitons of low energy might not escape branes, or might not get back in (it might polarize a brane and bounce off). AFAIK, we can't measure that. So 5 hidden branes might actually not give the 5:1 gravitational pull that we see.
A: All true. The simplest form of PDM ignores that, so it gives just a first approximation. To a second approximation, it might be one of the following:

• The required energy to escape/enter is so low as to be negligible.
• It's not negligible and there are more than 6 branes.
• The substrate is something other than branes.

If PDM comes to seem likely, it might let us see which is the case and provide a means for estimating the escape and entry energy.

So where's parallel earth?

Q: Where are the 5 parallel earths?
A: There aren't any. PDM does not propose that the other branes have parallel structure in their details. The neighborhood of our solar system is likely to be about as empty of dark matter as some randomly chosen equal-sized region of interstellar space.

How is it a halo?

Q: Why hasn't the dark matter collapsed into a galactic plane? Why is it still a spherical halo?
A: It will sometimes have collapsed individually for each brane but still in aggregate appear as a halo. Like a honeycomb paper party ball, each of its planes is flat, while the whole still approximates a sphere.

Of course the illustration overstates the case. It has a few dozen planes instead of 5 (dark), and their alignment is regular and efficiently makes an approximate sphere. In PDM the orientation of galactic planes should be random, so it doesn't make a sphere as efficiently.
Dark galaxies would probably be concentric with each other and with the associated visible galaxy, but random in orientation.

But that's MACHOs!

Q: But MACHOs were ruled out by lensing observations. That would rule out "parallel brane" MACHOs too.
A: The non-dark universe is mostly not massive compact objects. It's mostly neutrinos, and a lot of the rest is gas. And the MACHO observations leave room for a fair fraction of dark matter to be MACHOs, just not all or most of it.

How could it be tested?

Some predictions of PDM that seem testable:

• In a small percentage of spiral galaxies, the dark matter should be not a halo but a disk. I said earlier that, for dark spiral galaxies, their alignment is random with respect to our own and each other. So sometimes by pure chance they will all nearly line up.
  • Further prediction: such disks will have no particular orientation with respect to the visible galaxy's disk.
• Unseen stellar partners should be more common closer to the galatic center. That's because dark objects should be gravitationally capturable just like visible objects. However, they would tend to move relatively faster with respect to visible objects, because they tend to be orbiting in a different plane, so they're still individually less likely to be captured than visible objects.
  • The incidence of unseen stellar partners, in proportion to captured visible stellar partners, should be noticeably higher inside the galactic bulge, where all the dark companion galaxies overlap ours.
  • It should not vary so much within the galactic bulge.
  • It should have a very fuzzy boundary due to drift and to inexact overlap.
• Unfortunately I've lost the reference for this, so it's going to be vague. I figured I'd still put it out there as a possible disconfirmation. A few years back, astronomers found a flow of dark matter in our own galaxy. If PDM is true, it seems likely that such a flow would be part of a dark galaxy galactic plane. It should be approximately in a plane that goes thru the center of our galaxy.
  • There would be other dark flows in the same direction along the same galatic radius, within a few degrees to account for the dark disk's thickness.
  • There would be dark flows in the opposite direction along the opposite galatic radius, again within a few degrees.

General disclaimer

This is just a hypothesis I'm putting out there. If it's wrong, so be it.

Trade Logic: Branching quantifiers 2

Branching quantifiers in Trade Logic, part 2

Previously

Earlier I introduced Trade Logic. It's a form of logic that I designed to be suitable for connecting prediction markets. Last time, I built the branching quantifier Q_H in terms of Trade Logic, and now I'm going to build Q_L and Q_>=N

Q_L

Q_L (the Rescher quantifier) is a branching quantifier that basically means "less than or equal in number". That's one of the neat things that branching quantifiers can do that the classical quantifiers \(\forall\) and \(\exists\) can't.

As the description implies, we're going to need an equivalence predicate to build Q_L. Trade Logic deliberately doesn't provide a global equivalence predicate, so that will be a parameter.

For convenience, I'm also going to define iff (if and only if) as:

(or (and p q)(and (not p)(not q)))

It's not a predicate, because it takes formulas which are not an exposed type of object. With iff and Q_H, I can easily build Q_L. This is really just a transcription of the usual definition.

(define-predicate (Q_L 
                     U 
                     (the ,(pred thing thing) Equal) 
                     (the ,(pred thing) phi)
                     (the ,(pred thing) psi))
   (Q_H U
      (lambda (x_1 x_2 y_1 y_2)
         (and
            (iff (Equal x_1 x_2) (Equal y_1 y_2))
            (if (phi x_1) (psi y_1))))))

infinite aka Q_>=N

We can verbally define infinite sets as sets where no matter how many elements you enumerate, there is always another element. We'll treat that as an existential quantification outside a Q_L comparison. So I'll define infinite as:

(define-predicate (infinite U (the ,(pred thing thing) Equal) phi)
   (let
      ((p 
          (lambda (a)
             (Q_L U Equal phi
                (lambda (y) (and (not (Equal a y)) (phi y)))))))
      ;;"\exists a" around the entire Q_L part.
      (if (best p U +a) (and (phi a)(p a)))))

Side note: The LaTex formulas

I thought the LaTex formulas were formatting OK in Blogger because they looked OK in my previews. But apparently they're not. My mistake. I'll see if I can do something about them. But formatting math is not something I know a lot about.

Trade Logic: Branching quantifiers 1

Branching quantifiers in Trade Logic, part 1

Previously

Earlier I introduced Trade Logic. It's a form of logic that I designed to be suitable for connecting prediction markets. It is logic that "lives" within the system, not in individual traders' analyses.

As an experiment, I've been building towards writing the Internal Set Theory axioms¹ in Trade Logic formulas.

Branching quantifiers

The Internal Set Theory axioms require the predicate finite. This can be built using branching quantifiers. If we had used the "classical" quantifiers instead of best, we'd need to add more quantifiers now. Fortunately, branching quantifiers can be built with best, unfortunately not in a straightforward way.

We're also going to have to extend the mode system. Earlier, I said that the mode system required that there exist an ordering in which all variables are bound before they are used. But branching quantifiers bind variables simultaneously, in a sense. This sense is more obvious when using best, which requires applying the target predicate, which means that each branch needs values that are bound in the other branch.

We're going to start by building the simplest Henkin quantifier Q_H, then Q_L on top of that, then Q_>=N which means infinite. All will be parameterized on U (a universal set). They latter two also parameterized on an identity predicate.

Q_H

The basic idea is that the "test" clause of if is a conjunction of the best's of the two branches, while the formula dependent on all the objects is the "then" clause. In Trade Logic, if is always single-branched; it abbreviates (or (not A) B).

That almost works. The problem is that each branch actually nests two applications of best. We could easily rearrange it to get the outer variable, but we still need a way to retrieve the inner variable.

Each branch, if it were to exist in isolation, would be of the form $$ \forall x \exists y \phi(x,y) $$ which we write as²:

(let
   ;;Curried \exists y \phi(x,y)
   ((psi (lambda (X)
            (if
               (best (lambda (Y) (phi X Y)) U y_1)
               (Phi X y_1)))))
   ;;\forall x \psi
   (if (best (lambda (X) (not (psi X))) U x_1) (psi x_1)))

We'll define a slightly altered version of this that extract the variables, best_AE.

(define-predicate (best_AE phi U +x_out +y_out)
   (let
      ;;Curried \exists y \phi(x,y).  Still consults \phi
      ((psi (lambda (X)
               (if
                  (best (lambda (Y) (phi X Y)) U +y_out)
                  (phi X y_out)))))
      ;;\forall x \psi.  Doesn't consult \phi.
      (best (lambda (X) (not (psi X))) U +x_out)))

Note the resemblance between the formula for $$ \exists y \phi(x,y) $$ and a Skolemized variable. It takes a parameter, the same variable that $$ \forall x $$ binds. That means that when we incorporate this into a multi-branched configuration, that subformula will only "see" the universal binding that dominates it, which is just what we need for branching quantifiers.

With best_AE, we can define Q_H. NB, the branches of the conjunction output complementary arguments. First we'll define best_H:

(define-predicate (best_H Phi U +x_1 +x_2 +y_1 +y_2)
   (bind-unordered (x_1 x_2 y_1 y_2)
      (and 
         (best_AE
            (lambda (X Y)
               (Phi X x_2 Y y_2))
            U +x_1 +y_1)
         (best_AE 
            (lambda (X Y)
               (Phi x_1 X y_1 Y))
            U +x_2 +y_2))))

Sharp-eyed readers will notice two things:

• Contradictory ordering requirements. We have to bind x₁ before we can bind x₂, but also have to bind x₂ before we can bind x₁.
• New predicate bind-unordered. It works with the mode system and loosens the requirement that there exist an ordering in which all variables are bound before they are used. Within its scope, the ordering rules are suspended among the variables listed in the first argument. I'm not yet sure what the proper rules are for it. It's likely to block decomposition in most cases.

Having defined best_H, we can define Q_H as simply:

(define-predicate
   (Q_H Phi U)
   (if (best_H Phi U +x_1 +x_2 +y_1 +y_2)
      (Phi x_1 x_2 y_1 y_2)))

Footnotes:

¹ Why that? Just because I happen to like illimited numbers.

² I've changed pred to lambda. It means the same. As I wrote it more often, I realized that I preferred the familiar lambda even if it does wrongly suggest a function.

Trade logic: Objects are sampleable

All objects are sampleable

I mentioned earlier in passing that in trade logic, all things are sampleable. As I put it, "all variables are ultimately instantiatable".

Things can be finite sets of objects with some probability distribution over them. They can be trivially sampleable, such as single objects, where sampling them always gives the same object. It's even possible that they may measureable collections that have no distinct individuals, from which one selects measured extents.

And the time has come to nail down how Trade Logic is to do sampling (selection).

Population bets

Built-in predicate pick01

For a population to be sampled, we need some sort of random input, and it can't be vague where to get the random or how to use it. The purpose of pick01 is to provide the random parameters that sampling will use.

(pick01 +A) binds A to a real number between 0 and 1, lower-inclusive. Theoretically, A has infinite precision. In practice, precision should be provided lazily. There may need to be some provision against allowing requests for precision to tie up the system.

Settling population bets

I'll call a bet whose formula uses pick01 (explicitly or implicitly) a population bet. Population bets are useful for propositions that can easily be resolved for individuals but less easily for distributions.

Population bets that are trying to be resolved1 can pick particular individuals to resolve. Then they can settle probabilistically on the basis of that. It's basically using statistics.

The procedure for selecting an individual is to:

• independently choose a value for each pick01 directly or indirectly included in the proposition.
• Using those values, assess the issue's truth or falsity. (The Bettable selectors logic below should ensure that this step works)
• Repeat as wanted. We sample with replacement; for simplicity and generality, there is no provision for not looking at the same individual again.
• Using statistical logic (outside the scope here), settle the bet (maybe sometimes just partly settle it)

Selector bets

Of course randoms on the unit interval are not the only kind of random selection one might want. We have to build all sorts of selectors from them. At the same time, we want selection to be uniquely resolved. It won't do for a given selector to select more than one individual for the same randoms, or to select no individual at all.

So I introduce a different sort of bet: selector bets. They are bets that a given selector "works OK".

Any predicate that has only outputs can be the basis of a selector bet. In my current thinking, for each such bet, every output of the predicate would be associated with an equivalence predicate. That's needed so we have a principled way to distinguish individuals.

Unlike issues, the bet is asymmetric and its interpretation is adversarial: One side (challengers) tries to pick values for pick01 that they think give no result or too many results. The other side is betting that no such problems arise. I haven't fleshed out the exact rules for this.

best redux

It seems that I didn't adapt epsilon enough when I borrowed it as best. It doesn't make a sampleable output. Rather, its output is the set of all the most suitable objects in the universe. That's not neccessarily finite, much less measureable. So there's no good general distribution over that. It would be paradoxical if there were.

So best needs another parameter: a sampleable collection of objects. That can be any object in Trade Logic. This parameter will generally be quite boring; often it's simply the whole universe of discourse. Nevertheless, it's needed.

So =(best Pred From To)=² is now true just if:

• Pred is a unary predicate
• To is the part of From that satisfies Pred as well as anything else in From does

NB, I have not called To a "member" of From, as if From were a set. In general, From is a collection. To should be collection in general too. And we don't want to need an identity predicate as another parameter. And finally, we don't try to resolve "sub-part boundaries" here; those are addressed via structurally similar bets that replace calls of best with suitable particularizations. Nevertheless, I am not 100% sure I have supported collections adequately in this mechanism.

Footnotes:

¹ When a bet is "trying to be resolved" is beyond the scope here, but I have some ideas on the topic.

² I took the opportunity to put the arguments into in/out order.