Slashdot Mirror
Facebook Puts Deep Learning in the Palm of Your Hand (fortune.com)
Facebook has built a simple-looking video tool to show off a sophisticated use of artificial intelligence on cell phones. From a report on Fortune: During an event at its office in Menlo Park, Calif., last Friday afternoon, Facebook CTO Mike Schroepfer showed off software that takes a live Facebook video feed from a cell phone and converts the image in real time into a selection of artistic styles, such as that of Van Gogh. It might sound like a simple filter, but usually, an algorithm of this nature would need to send that type of information back to a server in a data center to process the pixels on more powerful machines. The Facebook crew crafted a less power-hungry and computing-intensive deep learning system they call "Caffe2Go," that uses the computing power in a cell phone. Facebook's Schroepfer showed the algorithm and other applications of artificial intelligence at the Web Summit conference in Lisbon, Portugal on Tuesday. Last Friday, he called the system a "pretty big leap" and "a real neural net running on a phone in real time."
What they're doing may be using machine learning technologies, but it isn't "deep learning", I speak as a person who's done a fair bit of work with machine learning. You can tell that their claim is bullshit right off the bat. Let me pull out the relevant bullshit quote
"It might sound like a simple filter, but usually, an algorithm of this nature would need to send that type of information back to a server in a data center to process the pixels on more powerful machines."
You see, this isn't true. Maybe facebook would usually send it back to more powerful machines, but those of us familiar with languages other than javascript would just write a simple C filter and do it locally in real time. I mean, my phone literally does it now. Like I can literally open up the camera app, switch it to video mode and select a filter. Machine learning involves statistical models and feeding new information into the original model to try to refine it. Machine learning is impressive with what it can do, but if you become familiar with it, you quickly learn what the limits are. Basically, if it can't be turned into a statistical model, then AI and machine learning are useless. Yes, applying a filter to an image to do what they're doing is typically done via a statistical model, which may say "AI" but it's also just saying "filter". And to reiterate, for machine learning, that statistical model has to be adaptive, as in it refines itself as more information is processed.