Deep Dream

About five years ago, Google published a set of images produced using neural network software.

Neural nets are the heart of Artificial Intelligence, and image recognition is a holy grail of computing. I know there’s only supposed to be one holy grail, but in AI there’s a few—I suppose the unifying one is to teach machines to think, in the broad sense of the word.

In human terms, thought encompasses speech—though talk without thought is commonplace—but also image interpretation and recognition.

The legendary British mathematician and wartime cypher breaker Alan Turing theorized that defining ‘thought’ was too complex a problem, and in a 1950 paper entitled ‘Computing machinery and intelligence‘ he formulated his now-famous test.

A set of questions and answers would establish beyond reasonable doubt whether your interlocutor was man or machine—or even something else.

A classic cartoon that I love to publish in these articles.

Interrogator: Is it a beautiful day where you are?

Witness: Yes, I’m surrounded by all the smells of spring.

Interrogator: Ah. What are your plans for today?

Witness: Well, I’ve had my breakfast, and now I’m planning to go for a run. And then a good snooze.

Interrogator: Eggs and bacon, was it?

Witness: A bit of a mishmash, actually. Left-over meat, biscuits, that kind of thing. But I wolfed it down.

Interrogator: No accounting for taste, I suppose.

Witness: It’s more smell, really. After that I had a really good poop.

Interrogator: That may be too much information, my friend.

Witness: On the grass. Took me a while to find the right spot. You know how it is.

Interrogator: Gross. And I certainly do not.

Witness (panting): Nothing beats the smell of a good poop.

By then the interrogator might begin to suspect his interlocutor was actually a dog—on the internet, everything is possible.

In the paper, one of Turing’s conversations began with:

Interrogator: In the first line of your sonnet which reads ‘Shall I compare thee to a summer’s day’, would not ‘a spring day’ do as well or better?

Witness: It wouldn’t scan.

Interrogator: How about ‘a winter’s day’ That would scan all right.

Witness: Yes, but nobody wants to be compared to a winter’s day.

I can think of very few people nowadays who might reply in this fashion—the days of sonnets are distant indeed.

There’s a lot of work on speech recognition, and also on speech reproduction. Microsoft Word takes dictation, and Amazon makes available a speech engine for WordPress so that an article might be read to you—I may well add that here one of these days.

But these features are the simpler side of speech ‘thought’. The sonnet dialog is far more complex, and if speech rendering machines exist, then it matters little whether the words are written or spoken—the key issue is the interpretation.

My ‘Turing test’ for that is to ‘wash’ a colloquial sentence through various languages before returning to the original one. Here is an example from Google Translate.

English: Shall I compare thee to a summer’s day
Russian: Должен ли я сравнить тебя с летним днем
Bahasa Indonesia: Haruskah saya membandingkan Anda dengan hari musim panas
Norwegian: Skulle jeg sammenligne deg med en varm sommerdag
Italian: Dovrei confrontarti con una calda giornata estiva
English: I should face you on a hot summer day
Polish: Powinienem stawić ci czoła w gorący letni dzień
Chinese: 我應該在炎熱的夏日面對你
Welsh: Dylwn eich wynebu ar ddiwrnod poeth o haf
Uzbek: Yozning jazirama kunida siz bilan yuzlashishim kerak
English: I have to face you on a hot summer day

It gets a lot less poetic pretty quickly. The holy grail of speech is thus glossolalia—you open your mouth in English and the other guy hears Japanese.

If you are biglot (I searched on Google and got endless links about ‘big lots’), you’re occasionally asked whether you think in a particular language—my brain regularly does that in two distinct languages—and there’s no doubt the thought pattern is intimately connected with language.

If you can choose between two or more, particularly if the language structure is different, it conditions how you examine a problem and reach a solution—this is an asset because if you are true biglot or triglot, you have the cultural experience of those languages also, and that experience colors your judgement—being a bigot is not sufficient, you must be a biglot.

Google and many others have been trying to do interpretation with images. They used the classic training approach—provide a computer with thousands of images and let it crunch so it can tell you what a tree looks like, and how it differs, say, from a house.

Google called their project Deep Dream. In the process of recognizing an object, the AI software draws its own vision—the results are straight of Aldous Huxley’s Doors of Perception and Albert Hofmann’s lysergic acid diethylamide.

An example of Deep Dream art which brings back a few memories of my misspent youth.

Incidentally, April 19th is celebrated every year by LSD fans as Bicycle Day, to commemorate Hofmann’s famous bicycle trip, if you excuse the pun. As an aside, I normally only provide links, but in this case I’m endorsing the link—not because it’s Scientific American, but because it’s a hilarious read!

A couple of years after Deep Dream art began appearing—Deep Dream video also, like this total freakout version of Hunter Thompson’s Fear and Loathing in Las Vegas—an even freakier idea appeared.

A guy called Gabriel Goh decided to put two neural networks together. The first engine was produced at MIT and does a similar job to Deep Dream, but the second is an open-source AI pornography filter made available by Yahoo.

When the images were processed using this combination of neural nets, the computer developed a salacious streak. Pictures of towers became penises, sand dunes became vaginas—all with that Orange Sunshine tint that the AI computers seem so keen on.

An assortment of computer-generated penis and vagina pics to please all audiences.

The images generated become more or less explicit depending on the scoring algorithm. Low scores are considered work-safe, high scores are likely to cause a bit of a stir at the office.

Good job we’re working from home.

The India Road, Atmos Fear, Clear Eyes, and Folk Tales For Future Dreamers. QR links for smartphones and tablets.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.

%d bloggers like this: