CNET también está disponible en español.

Ir a español

Don't show this again

Sci-Tech

AI valentine candy heart ideas: 'Stank love' and 'I honker'

What do you get after training a neural network to create romantic messages? Messages a lot fresher if sometimes more confusing than "be mine."

Today's artificial intelligence technology can find photos of our dogs and turn our speech into an email. Impressive. But is it creative enough to help us woo the objects of our affection on Valentine's Day?

That depends on whether you consider "hog" a term of endearment.

Janelle Shane, a research engineer for an optics company who also dabbles in neural network programming, trained a machine-learning system on about 360 messages found on valentine candy hearts, then asked the system to generate some new phrases to indulge our sweeties' sweet tooth.

AI valentine candy hearts

Research engineer Janelle Shane trained a neural network to create candy heart messages based on about 360 real-world messages. The results can be amusing.

Janelle Shane and Stephen Shankland/CNET

Some results sounded plausible -- "love bun" or "cute kiss," for example. Some of them are charmingly off kilter: "You are bare," "bog love" or "I honker."

The funniest ones, though, are the worst: "Stank love," "sweat pear," "you are boa" and of course "love 2000 hogs yea."

They may not pass quality control for the schmaltzy standards of the the greeting-card industry, but let's face it, they're a breath of fresh air. And they show how AI technology, even at this joking-around-hobbyist level, can find patterns we don't necessarily see.

"It puts the mirror back on us," Shane said. "It highlights the absurdity of what humans find attractive. Calling somebody a bug is something cute. Why not a hog?"

Machine learning today works by training a neural network -- a set of interconnected elements inspired by human brain nerve cells -- with source data like photographs, speech samples, or the text on candy hearts. Without understanding any conventional rules, the neural network can spot the kind of patterns that indicate there's a cat in a photo.

Neural networks for things like Google speech recognition take gargantuan amounts of computing power. But Shane's candy-heart experiment was a lot simpler. It took only about 10 minutes to train the neural network on her MacBook Pro.

"It took more time to go through the results than it took to train the neural network -- by a long shot," she said.

It's entirely up to us humans to decide whether results like "heat team," "wink bear" and "yak o way" should inspire grimaces at tech's shortcomings or admiration for a charming sort of innocence.

But hey, we're experts at anthropomorphizing clouds and cars. Why rule out something that's actually showing a glimmer of brains?