AI Is Learning to Meme
The Dango app uses deep learning to predict which emoji to use based on what you type.
The machines may be coming to take all of our jobs, but at least us meatbags can take solace in the fact that they'll never be able to replace our ability to find the perfect witty GIF to accompany an instant message... right?
Well, they kind of can—thanks to a new app called Dango launched on June 8, machines are starting to get pretty good at memeing, too.
Trying to teach machines how to naturally communicate with humans is a core goal of researchers working to create a general AI, or a machine that possesses the same level of intelligence as a human. Since memes and emoji are increasingly a part of what we consider human language (an emoji was the Oxford Dictionary's word of the year in 2015, after all), it was only a matter of time before someone applied neural networks to emoji to teach machines how to use them.
If you've ever fed photos into Google's trippy DeepDream program or had Facebook automatically tag your friends in a photo, you've encountered a neural network. This is basically just a way of teaching a machine by feeding it massive amounts of information from which it can abstract general data points until it starts to get an idea of what you're teaching it to recognize. In other words, feed the computer enough pictures of dogs, and it will get a general idea of "dog-ness," or what attributes to look for in photos to determine whether or not they contain a dog.
This is essentially what went into creating Dango, a floating app that will work on any messenger you have on your phone and help recommend emoji or memes that best match the message you're trying to convey. This is more complicated than it sounds—Dango isn't just scanning your message looking for the word 'horse' and then recommending the horse emoji. Rather, it is also taking into account the meaning of your message and other relevant factors, such as the fact that the eggplant emoji has, well, other meanings besides 'eggplant.'
When Dango's creators taught the underlying neural network how to meme, they fed it hundreds of millions of examples of emoji being used by humans from around the web. The ideas and emoji used in these examples were mapped into what Dango's creators call a "semantic space," where similar ideas and emojis are grouped together. So when you feed Dango a sentence, it uses its neural network to parse through the ideas in the sentence, and plot their vectors in this semantic space.
Once this is done, it "projects" the 1624 official Unicode Consortium emoji (which have also been mapped according to their connotations) onto this semantic space to determine where the textual and emoji vector points overlap. The points of overlap are determined to be the most appropriate emoji based on the semantic content of the text, and those are what get recommended.
Based on the massive dataset that went into making Dango, this means that not only is the app pretty good at 'understanding' the content of your message, it's also pretty up to date on pop culture. For instance, if you type in 'Donald Trump,' Dango will recommend emoji such as a man throwing away trash or a ballot tick. Type in 'Kanye West is the' and it will recommend a goat emoji, referencing Kanye's self-bestowed title of G.O.A.T (or, Greatest Of All Time).
So far, Dango has only been released in the Google Play Store, but if you want to give it a shot, you can try it out here.