With every year, machines surpass humans in more and more activities we once thought only we were capable of. Today’s computers can beat us in complex board games, transcribe speech in dozens of languages, and instantly identify almost any object. But the robots of tomorrow may go futher by learning to figure out what we’re feeling. And why does that matter? Because if machines and the people who run them can accurately read our emotional states, they may be able to assist us or manipulate us at unprecedented scales.
But before we get there, how can something so complex as emotion be converted into mere numbers, the only language machines understand? Essentially the same way our own brains interpret emotions, by learning how to spot them. American psychologist Paul Ekman identified certain universal emotions whose visual cues are understood the same way across cultures. For example, an image of a smile signals joy to modern urban dwellers and aboriginal tribesmen alike. And according to Ekman, anger, disgust, fear, joy, sadness, and surprise are equally recognizable.
As it turns out, computers are rapidly getting better at image recognition thanks to machine learning algorithms, such as neural networks. These consist of artificial nodes that mimic our biological neurons by forming connections and exchanging information. To train the network, sample inputs pre-classified into different categories, such as photos marked happy or sad, are fed into the system. The network then learns to classify those samples by adjusting the relative weights assigned to particular features.
The more training data it’s given, the better the algorithm becomes at correctly identifying new images. This is similar to our own brains, which learn from previous experiences to shape how new stimuli are processed. Recognition algorithms aren’t just limited to facial expressions. Our emotions manifest in many ways. There’s body language and vocal tone, changes in heart rate, complexion, and skin temperature, or even word frequency and sentence structure in our writing.
You might think that training neural networks to recognize these would be a long and complicated task until you realize just how much data is out there, and how quickly modern computers can process it. >From social media posts, uploaded photos and videos, and phone recordings, to heat-sensitive security cameras and wearables that monitor physiological signs, the big question is not how to collect enough data, but what we’re going to do with it. There are plenty of beneficial uses for computerized emotion recognition. Robots using algorithms to identify facial expressions can help children learn or provide lonely people with a sense of companionship. Social media companies are considering using algorithms to help prevent suicides by flagging posts that contain specific words or phrases.
And emotion recognition software can help treat mental disorders or even provide people with low-cost automated psychotherapy. Despite the potential benefits, the prospect of a massive network automatically scanning our photos, communications, and physiological signs is also quite disturbing. What are the implications for our privacy when such impersonal systems are used by corporations to exploit our emotions through advertising?
And what becomes of our rights if authorities think they can identify the people likely to commit crimes before they even make a conscious decision to act? Robots currently have a long way to go in distinguishing emotional nuances, like irony, and scales of emotions, just how happy or sad someone is. Nonetheless, they may eventually be able to accurately read our emotions and respond to them. Whether they can empathize with our fear of unwanted intrusion, however, that’s another story.