Microsoft Uses Machine Learning To ‘Read’ Your Face, Determine Your Emotional State

By now, it should be clear that machine-learning is turning into big business for the biggest corporations. We learned back in September that Apple is on a hiring spree to bring in some good talent, and Amazon even leases out the fruits of its efforts for researchers to take advantage of. Even more recently, we learned that Google loves and trusts its TensorFlow machine-learning library so much, it's willing to share it with the world by making it open source.

Machine-learning can be used for a multitude of things, both serious and not-so-serious. Microsoft has shown us in the past some good examples of the not-so-serious side, including an age-detection platform it showed off back in May which makes use of its Azure cloud platform. Now, the company has gone a slightly different route and is attempting to guess emotions.

Detecting emotions isn't new, but it stands to reason that as time goes on, it'll become far more accurate. Way back in 2007, a company called OMRON released technology that could pretty accurately guess the intensity of someone's smile. Microsoft's solution takes things a step further, by rating someone's face based on eight different metrics. It's worth noting that just this past summer, a company called Affectiva showed off very similar software; the results are a bit different, but both involve machine-learning and the same end goal.

Nonetheless, in the example seen above, Microsoft's algorithm detected that without a benefit of a doubt, this little girl is completely happy. Variations can occur in some readings, though, such as with the man at the absolute right of the photo. His happiness is 0.99797, while minuscule amounts of other emotions were detected (you'll have to click the source link below to see the readings for each individual in the picture). Humans immediately know that this man is completely happy, but machines go on by only what they know.

Since Microsoft doesn't want to depress us, none of the supplied examples show anything but happy people. However, this is a live demo, so you're able to import your own images if you want to, and let the detection engine do its thing. In quick tests with the help of Google's image search, we found the detection to be very accurate. The only metric we couldn't get to be detected was "contempt"; all others emotions were easily identified.