Microsoft at its Build conference showcased its progress in the field of machine learning with its new face detection APIs. Microsoft has put out up a demonstration of these APIs in the form of the How Old website, which uses the machine learning technology to detect the age of a user from a photograph.
On the website, users would have to upload their images in which the face is clearly visible and then wait for a few seconds for Microsoft’s engine to detect the age. Notably, the tool returns results for more than one person seen in an image. However, the results might not be satisfying as several users, including myself, were given dramatically incorrect estimates of their age.
The website duly notes that “Sorry if we didn’t quite get the age and gender right – we are still improving the feature.” Few other users however have also experimented with the company’s Face Detection API by uploading images of zombies, plastic dolls, and even emoji characters.
Being a zombie really ages you. http://t.co/dm3VwRHmwJ#Build2015pic.twitter.com/Hmpfv6OaZD
— Pete Pachal (@petepachal) April 30, 2015
OK Microsoft is straight up trolling my sister and the plastic baby here http://t.co/A3cpV0qszQpic.twitter.com/BKsuvz0Zkb
— Tom Warren (@tomwarren) April 30, 2015
— Hassan Khan (@hassankhan) April 30, 2015
Microsoft initially started the website as a test but within a short span of time it became popular over the internet. Microsoft’s Machine Learning Engineers Corom Thompson and Santosh Balasubramanian mentioned in a blog post that initially they were expecting about 50 users, but it went up to over 35,000 users within a few hours.
“We sent email to a group of several hundred people asking them to try the page for a few minutes and give us feedback – optimistically hoping that at least 50 people would give it a shot. We monitored our real time analytics dashboard to track usage and, within a few minutes, the number of people using the site vastly exceeded the number of people we had sent our email to,” they added .
The engineers said that the team also gets real time insights as and when users try the website. “For instance, we assumed that folks would not want to upload their own pictures but would prefer to select from pre-canned images such as what they found online. But we what we found out was that over half the pictures analysed were of people who had uploaded their own images. We used this insight to improve the user experience and did some additional testing around image uploads from mobile devices,” they added.