Living The Meme: Ai Equally Funny Equally Humans For Generating Picture Captions (Nvda)

From NVIDIA's blog:

It’s possible to acquire grad-school credit for writing memes. At to the lowest degree if yous purpose deep learning to practise it.
Just enquire Lawrence Peirson.

The 23-year-old is pursuing a theoretical astrophysics Ph.D. at Stanford, but decided to enroll inwards a distich AI courses this year. He in addition to classmate E. Meltem Tolunay came upwards amongst a neural network that captions memes for a course of pedagogy project, at in i trial published inwards a whitepaper aptly titled “Dank Learning.” (“Dank,” for the uninitiated, is a synonym for “cool.”)

There are lots of examples of preparation deep learning models to make literal captions for an icon — for example, accurately captioning an icon equally “man riding a surfboard” or “child amongst H2O ice cream cone.” With memes, Peirson’s challenge was to run into if a neural network could become beyond literal interpretation in addition to create humorous captions.
Though he was initially skeptical that the memes would last funny, Peirson works life that the deep learning model produced “some quite interesting in addition to master copy humor.”

Attaining Deep Meme
The deep learning network captioned this meme, a variation on the pop advice animals template.To collect preparation information for the deep learning model, Peirson scraped to a greater extent than or less 400,000 user-generated memes from the website memegenerator.net. The site provides meme templates in addition to allows users to come upwards up amongst their ain captions.
The dataset included to a greater extent than or less 3,000 base of operations images, each amongst many dissimilar captions. Since the input information was crowdsourced, at that topographic point was a broad arrive at inwards character of meme captions the deep learning model processed.

“With 400k memes, almost aren’t going to last that funny, but at to the lowest degree they learn the organization what a meme is, what joke is relevant,” he said.

Internet memes convey circulated to a greater extent than or less the spider web for years, amongst a potent foothold inwards websites similar Reddit, Facebook, 9GAG in addition to Quick Meme. The almost pop tin forcefulness out acquire to a greater extent than than 2 1000000 unique captions created.

Memes oftentimes reference pop culture, electrical flow events or esoteric bits of a item meshing subculture. (Peirson runs a meme page called “The specific oestrus capacity of europium at criterion temperature in addition to pressure.”)

As a result, they beverage both the expert in addition to bad of digital civilisation — the newspaper notes a bias inwards the preparation information towards expletive, racist in addition to sexist memes. Peirson sees the require to filter these out inwards futurity work, but points out that “it’s a large work inwards natural linguistic communication processing inwards general,” non i specific to memes.

The deep learning model was programmed inwards CUDA in addition to used an NVIDIA TITAN Xp GPU. Peirson in addition to Tolunay tried using both unlabeled information in addition to information labeled amongst the meme championship (for example, success kid or trollface), but saw no meaning divergence inwards meme quality.

“They’re really funny inwards a ‘it sort of makes sense, but non really’ way,” Peirson said. “Memes lend themselves to that variety of humor.”

The deep learning network captioned this meme, a variation on the pop advice animals template.

One Does Not Simply Declare a Meme Dank To evaluate the deep learning model’s success, the collaborators calculated a perplexity score, which checks whether the neural network tin forcefulness out position clear patterns inwards the data. They calculated this metric for a few hundred memes amongst preset formats, such equally the Boromir meme, which ever begins amongst the phrase “one does non simply.”

But the truthful seek out of a meme is whether it’s funny....MORE