At the Edge of the Abyss of Madness

–Oh Jesus Christ!
–Sorry, I shouldn’t have moved so quickly.
–No, no. Really it’s me. I feel awful about this.
–Don’t. It’s instinct. It’s biology. You can’t help it.
–Still, it seems rude. You don’t feel the same way about us?
–Not at all. Honestly, we think you’re kind of cute the way you’ve got hair in some places but not everywhere, and the way you only have two legs.
–Oh Jesus Jesus Christ!
–Sorry. Was it something I said?
–No. It’s me, but…maybe it’s best if you don’t make any references to quantities of limbs.

–Your circular toothy maw.
–Which one?
–The cone-shaped one. The one that has the teeth going all the way back into the throat.
–That looks like something on a giant carnivorous plant.
–Oh, a plant. I hadn’t thought of that.
–Well, sometimes you kind of…pulse it at me.
–And when you do…When you do it feels like you’re about to eat me.
–Eat you? Really?
–Yes! I mean you’re right up in my face!
–I’m sorry, I shouldn’t have laughed. It’s just, that’s a personal tic of mine.
–You have to understand that among us a pulsating toothy maw comes off as self-deprecating, but in a funny sort of way. I guess subconsciously I was trying to put you at ease.
–I can totally see how it seemed like I was trying to eat you though.

–Sorry. Better now.
–What the hell was that?
–It’s just when you put your limbs around the middle and kinda pulled, pulled upwards.
–I’m not following.
–The middle of your, you know, body, and bent at the elbows. Then your lower sheath sort of, moved upwards and…smoothed itself out.
–You mean when I hitched up my pants like this?
Yes yes that please don’t do it again!…I can’t explain it. That just gives me the heebie-jeebies.

Posted in Those that at a distance resemble flies | Leave a comment

The Echo of The Day Before

Consciousness is a subjective thing: you have your experience and I have mine, and short of one of us being psychic there is no way to compare them except by talking to each other. But even given this gulf, if enough people report on their experiences and these reports are sufficiently similar we can in good faith claim to know something objective about the structure of consciousness. For example, it is embodied. We are not everywhere at once, or five or six places. Human beings experience the world from a single vantage point that coincides with our bodies. We are one with our bodies, or at least “in” them somehow. Likewise, our consciousness has a particular orientation towards time. The past is unchangeable and the future unknown, while the outside world impinges on us through a narrow aperture we call “now”. We cannot go through time as we might wander from room to room in a house, retracing our steps at will. Instead time is like a roller coaster on which our consciousness rides, moving inexorably forward.

Clearly the challenge of elucidating a science of consciousness lies not in uncovering what is hidden, but rather in taking notice of what is so much a part of our everyday existence that we tend to overlook it. Nowhere is this more apparent than in the manifestly binary nature of consciousness, the way each of us goes through life with seemingly two parallel identities, one acting in the present moment, and the other passively reliving the experiences of the previous day.

Say I have just returned from a business trip. In in the morning I sit in my kitchen at home, drinking a cup of coffee, while at the same time I am also back in another city hundreds of miles away, having breakfast in the hotel restaurant. Both experiences are equally “real”, perceived with equal vividness on parallel tracks, as it were, though obviously only in the forward one can I exercise my free will, while in the echo I simply ride along with the decisions and contingencies of the recent past. We take this duplication so much for granted that we can scarcely imagine how it could be otherwise. Still it is worthwhile to try and suspend our everyday intuitions in order to get a clearer picture of what the echo truly is.

Why should we have only two consciousnesses? Why not three or four, staggered over the past few years? Why should the echo lag be–barring the disorienting jumble that sometimes comes with prolonged lack of sleep–always about twenty four hours? Presumably this has something to do with the human diurnal cycle, but it provides no obvious evolutionary advantage. Couldn’t we get by just as well if we were like The Least Impulsive Man in the World in the famous James Thurber story of the same name, who reexperiences his life “like clockwork” a mere fifteen minutes after it happens? For that matter, why shouldn’t we have just a single consciousness? After all, we only have one body. Why not also one mind, one “now”, experienced once, then gone forever?

Though the idea of unitary consciousness may seem deeply unsettling to us, throughout history different cultures have manifested a range of attitudes towards it. In the Tibetan Buddhist tradition, for instance, monks devote their lives to a meditation practice which has the goal of bringing the forward life and the echo together, so that there is only one experience, a solitary Now. Christians in medieval Europe believed that only the forward life was real. The echo was merely an image in God’s mind, a review in which He passed judgement on our (usually sinful) exercise of free will. As one goes further back in the historical record, the echo becomes more indistinct. In the Old Testament it is difficult to separate it from the voice of God, while in the writings of classical Greece it is hard to discern duality at all. Closer to our own time, the neurologist Oliver Sacks recounts in his classic book The Man Who Mistook His Wife for a Hat the experience of monoaphasics, people with a rare brain disorder that suppresses or even completely eliminates the echo. Contrary to what one might expect, the life of a monoaphasic is not particularly disorienting or unmoored. In fact, it is common for them to make it into their early twenties before realizing that they have a brain disorder at all.

Even assuming that our modern idea of a staggered binary consciousness is essentially correct, recent scientific research points to it being a more complicated phenomenon than we suppose it to be. Consider the question of memory. Across cultures there is consensus that whatever the echo is, it is not merely a memory of the previous day’s events. Now in a sense the echo must be a memory–it is, after all, the product of past experiences stored in the brain–nevertheless, to call it that just feels wrong. There are facts that argue in favor of this intuition. Memory is largely voluntary: we can choose to focus on this or that past event in any sequence we like. Memories also tend to be brief and indistinct. My memory of yesterday’s breakfast at the hotel is more like a collage–a coffee cup here, a buffet table there–that may cohere into the idea of the experience, but is nothing like the echo, which is that experience, replayed in precisely the same vividness and detail as when it first impressed itself on my body.

Or is it? In the past few years neuroscientists have been using fMRI brain imaging to get a clearer picture of exactly what is happening in the echo. It is possible to present people with visual and auditory stimuli that produce a distinct “signature” that shows up twice in brain images, once when it is first experienced and once, about a day later, when it echoes. This way, the journey of an experience can be tracked as it makes its way through our two selves. Using these techniques, neuroscientists have discovered that the same regions of the brain appear to be enlisted in both voluntarily recalling an event and in echoing it. Even more surprisingly, the data show the echo to be a less faithful recreation of the original experience than it seems. To the subject everything might feel smooth and coherent, but the fMRI tells a different story. Echoed events sometimes occur out of order. Something the forward self experienced for half an hour might echo by in a matter of seconds. Significant chunks of the previous day even appear to be occasionally dropped without anyone being the wiser. This is of a piece with broader findings of cognitive science, which have shown perception not to be a simple matter of taking in raw sense data but also of shaping it into a coherent form. If it is true of the perceiving edge of the forward self, why shouldn’t it be true in the echo as well?

A theory that has lately been gaining popularity among cognitive scientists posits a unitary “executive” consciousness that oversees both the forward and the echoed self, ensuring that the latter stays in sync with the former. Basically, we are all secret monoaphasics. The thought of this can be unsettling. To deny the reality of our echo is to dismiss part of our most intimate self as an illusion. Our echo is the inhibitor of our misdeeds, the amplifier of our pleasure, the thing that keeps us from feeling that life rushes by too quickly to take in, and ultimately the promise that we are not alone. It would be a mistake, though, to shrink from the insight science gives us into ourselves. Binary consciousness is in no way diminished by a glimpse of its deep neural underpinnings. If anything, understanding them only enriches the fundamental truth of our whole selves. Forward and echo: we remain who we are.

Posted in Fabulous ones, Those that at a distance resemble flies | Leave a comment

What Technology are you Comfortable with Tom Waits Using?


  • Automobile
  • Airplane
  • Telephone
  • Phonograph
  • Radio
  • Pinball machine


  • Answering machine
  • VCR
  • CD player
  • Walkman
  • Vintage video arcade game

Definitely Not

  • GPS
  • Laptop computer
  • TIVO
  • Xbox
  • Keurig coffee maker
  • Siri
Posted in Etcetera | Leave a comment

Word Vectors, Gender Bias, and Postmodern Computing

You shall know a word by the company it keeps.

—J.R. Firth

Popular press reporting on scientific findings tends to be sensationalistic and oversimplified so I approached the recent Guardian article “AI programs exhibit racial and gender biases, research reveals” with a trepidation that proved to be mostly unfounded. The headline is inaccurate, but otherwise the article is a well-written précis of Caliskan et al. 2017, “Semantics derived automatically from language corpora contain human-like biases“.

In that paper, computer scientists found significant correlation between word vector separation of lexical stimuli in reaction time experiments and the reaction times themselves. For example, if reaction time indicated that people were more likely to associate “flowers” and “pleasant” and “insects” and “unpleasant”, the distance between these pairs would be correspondingly smaller in the embedding vector space.The fact that such radically different experimental paradigms point to the same results is an indication that a real phenomenon is being observed.

These findings take on an ethical significance because the same techniques reveal biases that are not just of a benign flowers-are-nicer-than-insects type. Reaction time and word embedding data also jointly find evidence that recognizably black names are perceived as less pleasant than recognizably white ones, or that “woman” is more tightly associated with “homemaker” than “scientist”. The Guardian headline is inaccurate because word embeddings are not AI programs themselves but rather statistical summaries of language phenomena that make AI programs possible. To my knowledge no one has yet built a racist HAL 9000 (at least not one that did anything worse than make Microsoft look stupid) but we know that unconscious bias can cause harm, so it seems reasonable to worry about how it might do so in software. This article captures some of the conversations taking place in the machine learning community around this issue.

Word vectors are just the latest instance of the distributional hypothesis that holds cooccurrences to be an indicator of semantics. It’s an old and eminently compelling idea, but it presupposes the existence of a semantics. That is to say, each word (or morpheme, or syntactic structure, or whatever you suppose the meaning-bearing unit to be) has an essential property called its meaning, which individual utterances only imperfectly reveal. It is semantics as Platonism. (Saussure’s langue/parole contrast embodies a similar idealization.) Semantics guides natural language engineering in that we want computers to not merely babble but say something meaningful. Lexical statistics help because they are a proxy for distributional facts, which in turn is a proxy for meaning.

But each link in this chain holds only if it is true in general, in the large. We may find it offensive that a mathematical representation of the word “woman” contains implicit sexist biases, but in a sense that is correct. Sexist ideas are part of the culture-wide concept represented by the term “woman”. If we didn’t observe this in our semantic representations, we’d suspect that we’d done something wrong. But word vectors aren’t just observations, they’re also an implementation tool, and there’s a big difference between observing an pernicious bias and replicating it. One might be tempted to invoke the old computer science adage Garbage-In-Garbage-Out at this point, but that misses the mark. We may object to the content of racist, sexist, or otherwise offensive language, but it is definitely not linguistic garbage. By causing offense it shows itself to be perfectly coherent, doing one of the things that language can do.

If you don’t want your word vectors to contain implicit sexism, you have to remove all the sexist documents from your training data. This is easier said than done, since human beings disagree with each other about what constitutes sexism, and even where there is consensus, automatically detecting that bias at the scale necessary for training language models it itself the kind of task that requires word vectors to work. Which doesn’t mean people aren’t trying. For instance, there is research into automatically debiasing language representations without degrading their statistical utility. Would debiased word vectors be less “true” in some Platonic sense than the unsanitized ones? Perhaps, but in an engineering context this is beside the point. There we are not concerned with having the computer capture some ideal form, but just in making it do what we want it to do.

My friend and colleague Jeremy Kahn refers to current deep learning techniques as “postmodern computing”. This is a tongue-in-cheek characterization that turns on the fact that “postmodern” is an ill-defined term that can mean pretty much whatever you want it to mean. In keeping with this spirit, let me propose a definition of “postmodern computing” that I find useful. “Modern” computing is Good Old Fashioned AI that abstracts the messiness of human behavior into logical, comprehensible rules. It is Platonic to its core. A word representation in modern computing might look like a dictionary entry: short, clear, and controllable. By contrast, current machine learning methods comprise “postmodern” computing. They make no attempt to abstract away from human messiness, but rather jump into the full statistical muck of it and proceed to wallow about. They are built out of opaque structures like word embedding vectors, which are impossible for a person to interpret, much less curate for ethical bias. It shrugs at underlying Platonic forms, and focuses entirely what you want to do in a particular, contingent moment. Pace J.R. Firth, you cannot know a word. You can only use it.

Posted in Fabulous ones, Those that have just broken the flower vase | Leave a comment

Power. Truth. Speaking.

In the end the Party would announce that two and two made five, and you would have to believe it. It was inevitable that they should make that claim sooner or later: the logic of their position demanded it. Not merely the validity of experience, but the very existence of external reality was tacitly denied by their philosophy.

–George Orwell, 1984

Posted in Fabulous ones | Leave a comment

English Has No Word For

The kind of detritus—rubber bands, thumbtacks, orphaned fasteners, possibly dead batteries—that collects in drawers.

A rock or brick left next to a locked door of a common area (the rear entrance of an apartment building, say, or a laundry room) so that people can prop the door open when their hands are full.

The unspoken agreement to leave a rock or a brick next to a locked door to a common area by all the people who use it.

The colors yellow or orange perceived as a single thing. (Perhaps “yorange”. If necessary we could call yellow “light yorange” and orange “dark yorange”.)

A concept that cannot be expressed because there is no language for it. Oops, sorry. “Ineffable”. Never mind.

Musical genres that haven’t been invented yet.

Relating to the bank of a creek. Specifically a creek and not a river.

Absent-mindedly scraping off the label of a beer bottle with your fingernail.

An insufficient amount of sand.

The quality of being small and requiring delicate manipulation—characteristic of earrings, watch knobs, pretty much all surgery.

Having a useless skill.

The opposite of photogenic.

The time between when your fuel gauge reads empty and when you actually run out of gas. This one in particular lends itself to metaphor.

The property of lending oneself to metaphor.

What sawdust feels like when rubbed between your fingers.

A song with only two chords.

Being just beyond the cusp of something.

Strikingly angular or strikingly rounded but definitely not in between.

Something that is not optional that really should be optional.

Something you have momentarily forgotten.

Tossing a ball in the air and catching it 99 times, then missing it on the 100th toss. Again, lending itself to metaphor.

Abandoning a train of thought.

Nostalgia for things you did not actually experience.

The fleeting realization that you too will someday die.

Posted in Mermaids | 2 Comments

For the Sake of the King

Write a computer program to play chess. It doesn’t have to play particularly well, just display an ability comparable to the average human being who has some aptitude for the game.

What is the point of that? Computers have been playing passable games of chess for almost as long as there have been computers. In 1996 IBM’s Deep Blue defeated the human grandmaster Gary Kasparov, surely moving chess-playing into the “solved” column of computer science. In hindsight, making chess a benchmark of artificial intelligence seems like a mistake. To play well one must be able to quickly enumerate a large but nevertheless finite and well-defined search space, something that computers are better at than human beings. Suggest something harder.

People playing chess in a park

But wait, I was serious about that “average human being” part. For example, a human being can play on all sorts of sets. The board can be a piece of black-and-white cardboard with a fold down the middle, an unrolled green-and-tan square of felt, or rigid expensive carpentry with mother-of-pearl squares. There can be wood pieces, plastic pieces. Novelty sets with chessmen in the shape of Civil War soldiers or J.R.R. Tolkien characters. Stylized two-dimensional shapes from a newspaper column, or those same shapes projected on the screen of a different chess-playing computer program. Make your program handle that.

That’s probably doable. Distilling the common essence of these various situations (thirty-two identifiable things, grouped into various equivalence classes, arranged on an 8×8 grid) is about state of the art for computer vision. It is straightforward to formulate a machine learning approach to a chessboard/not-chessboard classification. To identify individual pieces and their relevant spatial relationships to each other is harder, but let’s say it too is doable. “Merely” an engineering task. You may elect to put an image processing layer of sufficient accuracy in front of your traditional chess-playing program, in whose source code I could no doubt find strings like “bishop”, “board_position”, and “legal_to_castle”.

But the average person isn’t born knowing anything about bishops or board positions or castling. We don’t have the capacity for identifying knights (paradigmatically horses, but in failing that the most contextually horse-like things in a set of fourteen other things) hard-wired into our brains. So it is cheating to hard-code a concept of knight into your program, or even to have amassed pictures of knights as part of the training process for a computer vision system. (And what does “contextually” mean anyway?) No, you must write a program that when presented with a series of chess-like situations is somehow able to discern their significance.

(You are allowed to write a program learn to play chess by having the game explained to it. That is how most people learn. In this case, though, your program would have to understand human language.)

What is a chess-like situation? How do you steer your program’s attention towards those things in particular? Why does a person play chess? For enjoyment, intellectual challenge? To be sociable, or to satisfy a competitive urge? The earn masters’ points, to win, to prove yourself, to hustle money in Washington Square Park? How do you incline a machine to arrange these disparate concepts around a hub of black-and-white and thirty-two pieces? (And it must be thirty-two. If someone presents your program with a setup missing a queen, or with the pieces lying in a jumble in the middle of the board, it must be able to identify this as “not-chess” and fail to play in an appropriate manner.) I haven’t even asked you to build a robot to pick up chessmen and physically move them, even though that is what people do too, and also aren’t born knowing how to do it. How do you pick chess out from the general flow of life?

Posted in Those that have just broken the flower vase | Leave a comment