Scientists generate memes using machine learning

You could have waited for it, but two scientists at Stanford University have developed a system that uses machine learning to generate memes that they say are indistinguishable from human ones.

This concerns the classic meme, ie an image with a ‘humorous’ subtitle and surname. The researchers wrote a Python script and pulled about 400,000 images from the site’s meme generator. They used this to train their model and to teach it which text belongs to which image. Then a so-called lstm network spat out an appropriate title on an image.

They used human judges to differentiate the product from real memes and to rate it based on its humor content. They think that memes have been rolled out that are indistinguishable from real ones. To allow other scientists and interested parties to build on their work, they have put the necessary tools on GitHub.

At the end of the research paper, they add: “Finally, we note that there was a bias in the dataset toward explicit, racist and sexist memes, so it’s an idea for future work to address this.” .”

Some results