What Happens When An Algorithm Interprets Your Halloween Costume

From Sony Betamax to...Sexy Beta Marx? JL

Janelle Shane reports in the New York Times:

Expect AI to try its best to give us exactly what we ask for, and be very careful what we ask for. It would be a bad idea to use the Halloween costume algorithm’s predicted costumes without checking them first (Spirit of Potatoes, anyone?). The algorithm learned to spell words by looking at costume examples. It makes predictions about which letters used in which order to make a costume, and then checks by looking at the data. If it’s wrong, it refines its internal structure. Its predictions get better. We know an algorithm can solve a problem, but we often don’t know exactly how.

When it comes to machine learning, things can get a little spooky: We know an algorithm can solve a problem, but we often don’t know exactly how.

Today’s machine-learning algorithms are considered a form of artificial intelligence, but it’s more helpful to think of them as prediction algorithms: Based on movies that this customer has rated highly, how much do we think he will like this other movie? Based on people we have hired in the past, how likely are we to hire this job candidate? Based on a list of past Halloween costumes, what might humans dress up as this year?

Hoping to find an answer to that last question, we turned to a machine-learning algorithm called textgenrnn that can learn to imitate text. Its author, Max Woolf, designed it as a blank slate ready to learn any kind of text; the text we gave it to imitate was a list of 7,182 costumes that people sent to aiweirdness.com over a year. Here are some examples of what that algorithm produced.

Zombie Schoolgirl

Toaster Boy

Donald McDonald

Zombie Schoolgirl

Donald McDonald

Toaster Boy

The algorithm learned to spell all of these words and phrases without human intervention, just by looking at the costume examples we gave it. It starts by making predictions about which letters should be used in which order to make a Halloween costume, and then it checks its own predictions by looking at the data used to train it. If it’s wrong (and at first, it almost always is), it refines its internal structure. Gradually, its predictions get better.

2. Comparing new

word with original data

3. Output and

structural refinement

1. Word generation

Deviled eggs

Ladybug

Mermaid

Witch

Wizard

Cinderella

T

A

I

B

B

W

C

B

C

H

Wabzb

U

I

L

Z

F

No original data match.

Structure is refined.

A

T

Q

R

P

Deviled eggs

Ladybug

Mermaid

Witch

Wizard

Cinderella

T

A

I

B

B

W

C

B

C

H

Witch

U

I

L

Z

F

Match found in data.

Structure updated to reflect

the success of this output.

A

T

Q

R

P

1. Word generation

T

A

I

B

B

T

A

I

B

B

W

C

B

C

H

W

C

B

C

H

U

U

I

L

Z

F

I

L

Z

F

A

T

Q

R

P

A

T

Q

R

P

2. Comparing new

word with original data

Wabzb

Witch

Deviled eggs

Ladybug

Mermaid

Witch

Wizard

Cinderella

Deviled eggs

Ladybug

Mermaid

Witch

Wizard

Cinderella

3. Output and

structural refinement

Match found in

data. Structure

updated to reflect

the success of

this output.

No original data

match. Structure

is refined.

Here are the costumes that the neural network produced at various stages, or “epochs,” of training. (We added the illustrations.)

Epoch 1

Ghanedastein

Heagd naller

Fustrastice

Vatand hampire

Captain Kirkf

Ruth Bader Hat Guy

Pirate the Wild Thor

Flip WItch

The Vampire Stab

Buttery Classic

Epoch 11

Sexy Beta Marx

Cat Witch

King Dog

Bucketball player

Dragon Ninja

baseball clown

Deatheater

Slick mermaid

Vampire Chick Shark

Centaur meije

When we ask it to predict new Halloween costumes, we can also tell it how creative to be. At the very lowest creativity setting it will almost always go with its top prediction, producing lots of repetition, while at the very highest setting it ventures into territory it deems less probable.

Sexy Santa

Creativity: 1

Note that at the beginning of training, the neural network knew nothing at all about its input data: It didn’t know the difference between letters and spaces, or how to spell any words. The more repetition it sees, the more easily it can learn something. Unsurprisingly, one of the first words it learns to spell reliably is “sexy.”

As its training progresses, it begins to learn more and more words, adding “steampunk,” “minion,” “cardinalfish” and “Bellatrix” to its vocabulary. It learns new and exciting ways to use “sexy,” including predicting sexy costumes that don’t yet exist:

sexy King Louis XVI

sexy michael cera

sexy printer

sexy marijuana bee

Sexy Tin Man

Sexy Minecraft Person

sexy abraham lincoln

sexy beet

sexy King Louis XVI

sexy michael cera

Sexy Tin Man

sexy abraham lincoln

sexy marijuana bee

sexy printer

Sexy Minecraft Person

sexy beet

The neural network also gets better at copying costumes directly from its training data. It does this because we told it to produce costumes like those it has already seen, and as far as the neural network is concerned, exactly copying its input data is a perfect solution.

Everything that the neural network knows about is something that was in its input dataset — it had no other information to work with. For example, it doesn’t know about this year’s hit movies, so it still produces “Pink Panther” more often than “Black Panther.” Even when it’s coining new words (as in “Farty Potter” or “Werefish”), it’s not building them based on any understanding of the meanings of the words but has simply determined that they are probable letter combinations based on the costumes it has seen.

Hello my name is

Get New Costume

And this is where neural networks and other machine-learning algorithms get a bit spooky: it’s very hard to tell how they determine that these letter combinations — like “Bloody Horse,” “Gothed pines” and “Ballerina trump” — are possible. Even when we can peer inside the neural network’s virtual brain and examine its virtual neurons, the rules it learns for its prediction-making are usually very hard to interpret. People are working on algorithmic explainability; for example, a group of Google and Carnegie Mellon researchers were recently able to show an image-recognition algorithm zeroing in on floppy ear shape as a way to recognize dogs. But many algorithms are much more difficult to interpret than image-recognition algorithms — for example, the kind used to make loan or parole decisions — and for the most part these algorithms are black boxes, producing predictions without explanation.

Sometimes, the ways algorithms work can have unexpected and disastrous consequences. In 2013, M.I.T. researchers trained an algorithm that was supposed to figure out how to sort a list of numbers. The humans told the algorithm that the goal was to reduce sorting errors, so the program deleted the list entirely, leaving zero sorting errors. And in 1997, another algorithm was supposed to learn to land an airplane on an aircraft carrier as gently as possible. Instead, it discovered that in its simulation it could land the plane with such huge force that the simulation couldn’t store the measurement, and would register zero force instead.

It can be a scary thing to trust a decision that we don’t understand — and it should be scary. Machine-learning algorithms learn from whatever is in their training data, even if their training data is full of the behaviors of biased humans. In other words, when we are using machine-learning algorithms, we get exactly what we ask for — for better or worse. For example, an algorithm that sees hiring decisions biased by race or gender will predict biased hiring decisions, and an algorithm that sees racial bias in parole decisions will learn to imitate this bias when making its own parole decisions. After all, we didn’t ask those algorithms what the best decision would have been. We only asked them to predict which decisions the humans in its training data would have made.

Landing a plane

Parole decisions

Job hiring

Landing a plane

Job hiring

Parole decisions

This kind of mistake happens all the time, and biased algorithms are everywhere. The moral of this story is not to expect artificial intelligence to be fair or impartial or to have the faintest clue about what our goals are. Instead, we should expect AI merely to try its best to give us exactly what we ask for, and we should be very careful what we ask for. It would have been a bad idea to put the plane-landing algorithm’s recommendations into practice without checking them carefully. And it would be a bad idea to use the Halloween costume algorithm’s predicted costumes without checking them first (Spirit of Potatoes, anyone?).