Facebook's Truly, Madly Deeply Learning the Story of Your Life
Facebook has set up an eight-person team to look into how artificial intelligence can help it further analyze data it gathers on its members, the MIT Technology Review reported.
The team will work with an emerging AI technique called "deep learning."
In a possibly related development, Facebook updated a patent filing for real-time content searching in social networks.
What Deep Learning Can Do
Deep learning can let software work out emotions or events not explicitly referenced in people's writing.
It can also let software recognize objects in photos -- and, based on the large amount of data analyzed, it can make sophisticated predictions about people's likely future behavior.
"Supermarkets spend a lot of money trying to figure out what products they should put next to the checkout register," Jim McGregor, principal analyst at Tirias Research, told TechNewsWorld.
"With Deep Learning, you can tell someone's favorite color and target them more closely," McGregor said. "If you like the color black, all the marketing materials I send you will be in black, for example."
Facebook reportedly will use the results of deep learning to streamline updates to members' newsfeeds, and help people organize and manage their photographs, among other things.
It might also use the results to better target ads.
What Is Deep Learning?
Deep learning uses a set of algorithms that try to learn layered models of inputs. The layers in such models correspond to different levels of concepts.
Higher-level concepts are defined from lower-level ones, along the lines of inductive reasoning.
Most of the deep learning algorithms can be applied to unlabeled data, even when the data cannot be associated with the immediate task.
In other words, like human minds, deep learning algorithms take in a multiplicity of stimuli or facts that are not necessarily related, and work with them to achieve some level of knowledge or understanding.
A multilayered neural network can create internal representations, and each layer can learn different features. For example, one layer could learn the orientations of lines; the second may combine these to identify simple shapes; the next layer may then create more abstract shapes; and finally, they could be put together to classify an image.
A multi-level hierarchy of recurrent neural networks can be trained through unsupervised learning one level at a time, and fine-tuned through the use of a backpropagation algorithm.
Backpropagation occurs when you get the desired result of an operation and, from that, figure out the inputs that gave you the result.
Implications of Facebook's Deep Learning
Opinions differ as to whether Facebook's plans to use AI to more thoroughly mine information members post on its site are troubling.
"I don't think this is much different from what they already do," Justin Brookman, director, consumer privacy at the Center for Democracy & Technology, told TechNewsWorld. "It's just a new way for them to organize and analyze data they already have."
Privacy advocates "are more concerned by the underlying data collection and retention than in how it's used for advertising," he continued. "Facebook might want to consider giving users an option to not have personalized ads at all."
On the other hand, "you can call it by many names, but it's basically data analysis on steroids," contended Tirias' McGregor.
"If you really get into it, you can identify someone from their pictures and their friends," he said. "This is what police departments and law enforcement [agencies] do -- develop profiles of people."
The danger is that "Facebook or anybody tied to Facebook can sell this data to anyone, so they'll be putting out a profile of you in the public view that they can sell," McGregor suggested.
Facebook did not respond to our request to comment for this story.