Google, Facebook, Amazon advance machine-learning applications

Teaching machines to read/comprehend websites, recognize and group faces, and reject fake reviews
June 22, 2015

Three new significant developments in machine-learning were announced last week.

Reading and comprehending natural-language documents

Google DeepMind in London said it has developed a way to teach machines to read natural-language documents and comprehend them, and like Watson, answer complex questions with minimal prior knowledge of language structure — at least for CNN and Daily Mail websites.

Daily Mail bullet points (credit: Daily Mail)

As noted by the researchers in an arXiv paper (open access), these websites have summaries (such as bulleted lists) and paraphrase sentences. The researchers were able to use these for creating context–query–answer triples for each document. In the process, they generated two new corpora (collections of data) of roughly a million news stories with associated queries to serve as training sets.

Facial recognition for sharing photos with friends

Facebook has launched Moments, an app that uses facial recognition technology to groups the photos on your phone based on when they were taken and, using facial recognition technology, which friends are in them. You can then privately sync those photos quickly and easily with specific friends, and they can choose to sync their photos with you as well.

Syncing photos to friend in Moments (credit: Facebook)

The app and this technology is based in part on work conducted by the Facebook AI Research (FAIR) team, headed by AI research Yann LeCun, as he explains in this video:


Facebook | Facebook AI research

But an experimental algorithm created by Facebook’s FAIR lab can recognize people in photographs even when it can’t see their faces. Instead it looks for other unique characteristics like your hairdo, clothing, body shape and pose, New Scientist notes.

“The research team pulled almost 40,000 public photos from Flickr — some of people with their full face clearly visible, and others where they were turned away – and ran them through a sophisticated neural network. The final algorithm was able to recognize individual people’s identities with 83 per cent accuracy. An algorithm like this could one day help power photo apps like Facebook’s Moments.

“LeCun also imagines such a tool would be useful for the privacy-conscious – alerting someone whenever a photo of themselves, however obscured, pops up on the internet. The flipside is also true: the ability to identify someone even when they are not looking at the camera raises some serious privacy implications.”

 Amazon machine learning algorithm fights fake product reviews

Amazon has developed a machine learning algorithm that will “learn which reviews are most helpful to customers” — that is, which reviews are real and which ones are fake. (Amazon sued a number of websites that specialized in creating fake Amazon reviews in April.) Amazon will give greater weight to newer, more helpful and verified customer reviews and ratings (their 5-star system).

Amazon Web Services began offering its Amazon Machine Learning  service in April to make “it easy for developers of all skill levels to use machine learning technology … without having to learn complex ML algorithms and technology.”