Facebook is Training its AI to Read Human Facial Expressions

Wednesday, 06 September 2017 - 8:27PM
Technology
Artificial Intelligence
Wednesday, 06 September 2017 - 8:27PM
Facebook is Training its AI to Read Human Facial Expressions
Pixabay
The artificial intelligence experts at Facebook are at it again. The social media juggernaut recently published an open paper on the challenges that have been faced in their new endeavor: teaching robots to smile without looking creepy.

Fair enough, that's a pretty important nut to crack, and should have far-reaching effects if Facebook's scientists and engineers can pull it off.

The paper details the process of teaching Facebook's developing AI babies to recognize human expressions when viewing faces. The focus here isn't on picking up on particular emotions such as joy or sadness, but rather, on learning to differentiate between subtle facial cues (often called "microexpressions") that all people make, and which we notice subconsciously without realizing it - and similarly notice when presented with a human facsimile that lacks these emotional ticks.

The aim, essentially, is to help future Facebook AI to avoid the dreaded Uncanny Valley by making simulated facial expressions better mimic the way that real humans emote when communicating. This is something that, traditionally, AI has been very bad at pulling off.

This isn't the AI's fault; because humans watch faces from essentially the moment we're born in an effort to study the emotions of others around us, this is one of the hardest things for a computer to fool us on. We're all subconscious experts at reading each other, and any seemingly imperceptible deviation from standard facial movement sets of warning sirens within our brains that something isn't quite right about the robot we're holding a casual conversation with.

If Facebook's engineers can use machine learning to crack the code behind microexpressions, then we'll quickly end up in a situation where simulated assistants or robots will feel more genuine, trustworthy, and alive. It would be a phenomenal step forward for human-robot relations in general, if only because people would stop being repulsed every time they see a robot trying desperately to mimic human behavior.

Naturally, of course, the idea that Facebook's AI is learning the ins and outs of how humans communicate nonverbally is a little worrying to robot naysayers, or anyone who takes Elon Musk's comments at face value. Considering that Facebook recently had to shut down an AI for developing its own secret language, there's a danger that this tech could help future artificial intelligences learn a little more about human behavior than we'd like.



It's hard not to think of Ex Machina here. The idea of robots learning our microexpressions represents our collective fears about what rogue AI could accomplish if they perfectly understand how to make humans trust them. Robots that have exposure to a wide enough sample of internet users will inevitably learn the social benefits of being cute when trying to get us to open up to them.

The good news is that, at present, the AIs don't have any specific reason to use their newfound knowledge of human nonverbal communication for evil. If anything, the bigger threat is Facebook using their new microexpression-reading AI to scan our photos and videos, in order to learn when we might be more susceptible to ice cream marketing.

Sure, your status says you're doing great since your break up, but your eyes tell a different story. Have a look at this sponsored post for Ben & Jerry's, etc.
Science
Science News
Technology
Artificial Intelligence
No