Religious Language

Two forms of language are commonly used to describe the infinite: via analogia and via negativa. Via analogia is used when the Christian says "God is my shepherd", while via negativa is used when the Buddhist says "the infinite is unchanging". These two forms are normally presented and discussed separately — I think there's a more elegant way of describing them: both can be seen as explanation by analogy. That is, both these forms describe something unfamiliar by the means of something familiar. Via analogia does this by comparing positive characteristics (similarities), while via negativa compares negative characteristics (differences). What's interesting is that the positive comparisons are more prone to comparing groups of characteristics while negative comparisons are more about single characteristics (we say, more often, "God is not x" rather than "God is not like x"). I expect there is a neural efficiency explanation for this bias, but it seems like the opposite bias we'd expect.

Native American Messiah

At the end of the 19th century, while white Americans were busy decimating the Buffalo population, some Native American shamans shared a vision of what is called the "Ghost Dance":

All Indians must dance, everywhere, keep on dancing. Pretty soon in next spring the Big Man come. He bring back all game of every kind. The game be thick everywhere. All dead Indians come back and live again. They all be strong just like young man, be young again. Old blind Indian see again and get young and have fine time. When Big Man comes this way, then all Indians go to mountains, high up away from whites. Whites can't hurt Indians then. Then while Indians way high up, big flood comes like water and all white people die, get drowned. After that water go way and then nobody but Indians everywhere and game all kinds thick. [...] Indians who don't dance, who don't believe in this word, will grow little, just about a foot high, and stay that way. Some of them will be turned into wood and be burned in fire.

There are some obvious parallels to the Christian notion of the Second Coming, but what I find more interesting are the differences. Through various tensions and misunderstandings, the movement was dealt its strongest blow at Wounded Knee in 1890 where the US Calvary killed at lest 153 Sioux. The Ghost Dance movement quickly fade in response to this defeat. When prophecies aren't fulfilled, shouldn't they be ignored? In this respect, the faith of the Native American seem more reasonable than Christianity (excluding extremely liberal interpretations of "Christianity").

Justified Subjective Belief

I met this guy recently, Eric, who prefers to make the distinction between being a Christian and a Paulian. As a Christian, he doesn't try and defend Paul's writings (or scripture in general), but stands by Christ's central teachings. It's questionable what Christ's central teachings are (and how original they are), but Eric says it's just about love. He sees this as objectively defensible, but it seems like the teaching is dependent on the character of Christ: if Jesus was just a person, believing in "Love" is not being "Christian", it's also being Buddhist, Confucianist, and a host of other things. On the other hand, if Jesus is God incarnate, He has a privileged understanding of the nature of humanity and the ethics that follow. Arguing for the latter requires a shareable (objectively defensible) argument for Christ's God-nature, which would have to be founded in scripture.

A more general question that arises from this regards the nature of personal belief. Eric believes more than he can share (again, objectively defend), but isn't worried about not being able to share it. This concerns me: when, if ever, are we justified in believing something about the nature of reality we can't defend to someone else?

Communication and Art

Communication that relates information can be classified on a spectrum from "poetic" to "prosaic". On the prosaic end we have things like chemistry catalogs, stock prices, and math text books, while somewhere in the middle are novels and certain photography, and to the deep poetic end we have abstract expressionism, some dance traditions, and zen koans.

Prose is characterized by its explicitness. If we imagine human knowledge as a collection of nodes of information, and connections describing the relationships between them, prose is used to introduce new nodes.

Poetry is characterized by its implicitness. Instead of introducing information directly, it is used to induce indirectly. It comes in two forms: adaptive poetry and explorative poetry. Adaptive poetry makes use of analogy to carry relational information from one situation to another (music, and a lot of common poetry fits this category). Explorative poetry forces the listener to apprehend the meaning directly for themselves. Perhaps by making an observation of the external world, or more often through reflection (as with koans).

Under this description of communication, art can be seen as synonymous with poetry. There are two dimensions in which art can further be described: whether it is received as art, and whether it is created as art. This allows for three subcategories:

• Created but not received as art: Failed Art — for example, elevator music (assumes the creator and receiver are not the same person)
• Received but not created as art: Ironic Art — for example, rainbows, clouds, microscope slides viewed by an artist, possibly some archaeological artifacts
• Created and received as art: Traditional Art — most paintings, written poetry, music, etc. (Distinguishing between the intended and received message would be counter productive.)

The Expansion Problem

Given an initial corpus of text for training and a target text, find the minimum number of letters the text can be reduced to while maintaining enough information to be reconstructed without ambiguity.

For example, the training text might be a few hundred articles from the New York Times. The target text might be this sentence. Perhaps we could encode the target text:

t targ tex mi b thi sent.

There encoding and decoding of these sentences are deeply connected. There are also multiple levels of information. The first two can be represented by a Markov chain algorithm:

1. Words: "thereb" encodes "thereby", because nothing else starts with "thereb"
2. Grammar: "tha man" encodes "that man" because "that" is most likely to precede "man" and start with "tha".
3. Context: More general than grammar, when talking about flowers we might talk about colors as well, etc. This is a long-distance relationship between words.

The problem has a huge search space, but might be efficiently implemented with some creative heuristics. The applications to natural language compression is obvious, but I think this would be nice in a real-time system that expanded your text as you typed (say, an email). For keyboard-based input, the slowdown in watching for text expansion would probably outweigh the benefits, but in situations where entering input is an expensive operation, this would be ideal (I think some cell phones implement a version of this that considers the first-level, word frequency, for key disambiguation). There are also creative applications if you apply this to music (a language, just generally limited to poetic communication).

Sitting Down

Last year I had a short conversation over lunch with an RPI girl. She was telling me how suffocated she felt:

I want to go on a safari. I want to go to a party and play games. I want to go to an art museum. I want to go shopping. I don't feel like I'm living, but always sitting down.

I was conflicted — it makes me glad to see passionate people, but the desire to go should be balanced with a contentness. I just now found a note I wrote to myself later that day:

That "real life" is "somewhere else" is an illusion... The desire for the unknown is echoed, but there is a failure to recognize the beauty of the present... We must live the present as if we are already in a far off land — because we are!

This joy of contentness coupled with hope is something I wish I could explain better. If I was given a few wishes, I would ask for the best way to explain this.

Update: a reminder from Emerson: "Art is not to be found by touring to Egypt, China, or Peru; if you cannot find it at your own door, you will never find it."

Dual Filesystem/Mail Server

System administrators do their best to discourage people from sending large files via email. It hogs resources that need to be used for quickly sending short messages, and mail clients aren't generally made to handle lots of large attachments. I think it would be wonderful to have a system that completely rethought the idea of email and shared file storage, and combined them with some of the ideas behind version control systems.

If we imagine everything as data and links, emails become text data with links to other files that are normally called "attachments". There are many different types of links — the metadata on the email (like to/from/subject etc.) would be separately stored and linked to using the respective link types (a "to" link, a "from" link, etc). Forwarding turns into the act of identifying a new recipient for an email and creating a link to them. Replying would entail a link to the replied-to email. "Recieving" an email would simply be recognition of a "to" link associated with your address. If someone would like to modify the document "attached" to an email, it would be edited locally and then added as a branch of the original version — ready for merging if others agree on the changes.

The strength of this would be in a corporate or academic environment. The difficulty comes in when you get people from the "outside" replying to your email with something that is hard to automatically recognize as a "reply", etc. Handling permissions is something else to think through.

Arboreal Proliferation

It's really no wonder that trees survive — there's so much surface area for them to soak up life. It's interesting to see the variety in trees, and how they all get along in their own ways. The big ones need more sun and water and nutrients, which they get because they're bigger — and the small ones don't need as much, but they still get it.

Intersecting Summation Compression

If we have the summation of each row and each column for an image, how much of the image can we reconstruct? There are obviously multiple solutions once you get beyond the 2x1 pixel image, but what if we add extra information about images in general — like the fact that two nearby pixels are similar colors in general? And if this isn't enough information, what if we have the diagonal summations as well? Multiple diagonals? Other angles besides 45 degrees? What does the shape of the linear summation information versus recall accuracy function look like?

Compression as a Learning Problem

Humans are pretty good at efficiently communicating information to each other. One of the reasons for this is that we can model the listener's expectations (their interpretation) of what we are saying. If I say "Can I have that?", pointing towards the table, you will hand me the pen because you can see I just pulled out my checkbook. We have long-term expectations, a learned pragmatics to conversation (like you knowing I want to write a check), and short-term expectations, like learning the meaning of pronouns.

If we apply this to a computational context, we get an interesting compression algorithm. Imagine computer A trying to send a file to computer B. It's a binary file, and because it's not noise, there are short patterns here and there. A and B both have the same predictive capabilities — let's say they're both using Markov models — so they can guess what the next bit is with a certain confidence. So here's what happens: A starts sending B bits, and B starts learning patterns in the bits being sent. A is modeling B's mind, so it knows what B is expecting. Now if A knows that B is very confident about the next bit, A doesn't even bother to send it, it just moves on. Thus you only transfer a portion of the information, and the rest is implied.

The problem, of course, is how expensive it is for A to correct B if it makes a wrong prediction. Bit-by-bit would have a lot of overhead, so it would probably be best for A to send a long sequence at a time, coupled with a note about any bits B guessed wrong.

Nested Links

When I was just starting to learn Polish, I'd have trouble remembering the pronunciation of some letters. For example, if I ran across the city Łódź, I had to look up the ł, ó, and ź to realize it was pronounced "woodge". It would be nice if there was a technique for nesting hyperlinks within each other, and then allowing the user to specify the level using their mouse position (the height — closer to the baseline would mean a deeper level). So if I wrote code that nested the name [[Ł][ó]d[ź]], it would allow me to click on any of the three letters, or the entire link.

Individual letters is probably a language-specific use, but this makes a lot of sense when linking to information composed of difficult words — each word could be associated with a dictionary entry, while the whole phrase would be the normal link. Or if you were looking through a geneaology, and you see "John Smith": maybe you want to follow that exact name, or just his family name, or find out about the frequency of his first name.

The Problem of Corruption

Imagine a jail filled with prisoners of different psychological makeup. They vary from one to another in many dimensions, but let's assume everyone can be rated on a single dimension: goodness. And we'll make one more assumption: good people inspire good in others, and evil people inspire evil in others (to the degree which they are good/evil). The problem: is it possible to organize the activities of the prisoners so that the overall goodness of the jail is increased over time?

If we imagine this situation as isomorphic to some other better-studied phenomena, say, heat distribution, the answer is obvious: there's no solution. Imagine each prisoner as a hot object (heat corresponds to goodness). there's no way of organizing the interaction of the objects to produce more heat than was initially available.

Some possible hopes for our prisoners:

• There may be some way of bringing outside sources (i.e., society at large) into jails to shift the curve for the best.
• Because psychology is more complex than a goodness value, perhaps the effect of interaction is not symmetric, and goodness inspires more goodness then evil inspires evil? If we add another dimension to the problem, say extroversion and introversion, maybe this would allow us to augment interaction beneficially?
• Perhaps the jail system can be subdivided into different tiers of goodness, and the best from each would slowly be moved to the next level? Group behavior may provide unexpected side effects.

Variable-Depth Markov Models

What would a Markov model act like if we allowed variable depth instead of fixing it? First, if you imagine the MM going top-down from initial to final states, it might make sense to group states that always appear in the same context — horizontal generalization into a state-type. Allowing variable depth could be implemented by vertical generalization into state-chain-types. In text processing/generation, this might appear as the creation of single states from idiomatic expressions.

Spamming Turing

One of the reasons I love Larry Kagan's art is for its origin: the shadows were originally a problem he sought to eradicate instead of a feature to take advantage of. What if we apply this idea to spam email? What could it possibly be good for?

If we look at the problem on a document-level, there is the potential for communicating relevant information in the form of spam. But we already know this doesn't work: nothing is relevant to everyone with an email address (much less to everyone in your address book, as some people have proven to me with their cute animals/national anthem/animated gif forwards).

On the other hand, if we look at it on a larger structural-level, we see something more interesting: at least half a billion email addresses receiving spam, most of which have some sort of spam filter in place (this is a guess based in the popularity of Yahoo! mail, Hotmail, Gmail, etc.). Content-based spam filters are, in a sense, fitness functions for the human-ness of a message. What's more, when an email gets past a spam filter, you get a real live human to decide whether its legitimate or not.

Ignoring any ethical dilemmas, I propose a learning system that makes an attempt to "reach out" to others via email, revising its attempts based on the clicks each different email receives (of course, there would be a URL in the message). I predict it will derive a shorter version of the Nigerian email scam, or something with the same theme ("I'm in need of trouble and need a response").

"En Það Besta Sem Guð Hefur Skapað..."

It seems like we have a hunger for "newness", or death and rebirth. In Christianity, we have the death of the "misdeeds of the body" (Romans 8:13) and rebirth in Christ (and in Judaism, the Jubilee year). In Buddhism, the total death of self and recreation (or perhaps "realization") of oneness with everything. Academia is colored religous by its semester-oriented structure; any student can explain the "fresh" feeling of a new semester. Sartre takes this to an extreme, saying that we are new creations every moment (which, oddly enough, causes angst). Total permanence, reminiscent of Parmenides, traps us. Total impermanence, a la Heraclitus ("Everything flows, nothing stands still."), frightens us and causes angst.

Perhaps the love of shopping that so characterizes our American society thrives on this dichotomy? Materialism is a continual recreation through addition, coupled with an eventual eradication of the old. (Being the only sort of "recreation" people know, it's not suprising that Christian speakers spend so much time discouraging this attitude in the context of Christianity.) Materialism provides a sense of newness without forcing us to identify with a purely permanent or impermanent nature.