Twitter: Short Text Outfit Gets Excited about Computer Vision

July 30, 2014

Robots. Magic stuff like semantic concept lenses. Logical wackiness like software that delivers knowledge.

I read “Twitter Acquires Deep Learning Startup Madbits.” The write up points out to the drooling venture capitalists that Twitter’s purchase is “the latest in a spate of deep learning and computer vision acquisitions that also includes Google, Yahoo, Dropbox, and Pinterest.” What this means is that these oh-so-hot outfits are purchasing companies that purport to have software that can figure stuff out.

I recall a demonstration in Japan in the late 1990s. I was giving some talks in Osaka and Tokyo. One of the groups I met showed me a software system that could process a picture and spit out what was in the picture. I remember that the system was able to analyze a photo of a black and white cow standing in a green pasture.

The software nailed it. The system displayed in Japanese, ?. My hosts explained that the idiograph meant “cow.” High fives ensued. On other pictures, the system did not perform particularly well.

Flash forward 30 years. In a demonstration of image recognition at an intelligence conference, the system worked like a champ on clear images that allowed the algorithm to identify key points, compute distances, and then scurry off to match the numeric values of one face with those in the demo system’s index. The system, after decades of effort and massive computational horsepower increases, was batting about .500.

The problem is that different pictures have different looks. When the subject is wearing a baseball cap, has grown a beard, or is simply laughing, the system does not do particularly well.

You can see how Google performs. Navigate to Google Images, select a picture of a monument, and let Google find matches. Some are spot on. I use a perfect match example in my lectures about open source intelligence tools. I have some snaps in my presentatio0n that do not work particularly well. Here’s an example of a Google baffler:


This is a stuffed pony wearing a hat. Photo was taken in Santiago, Chile at an outdoor flea market.

This is the match that Google returned:


Notice that there were no stuffed horses in the retrieved data set. The “noise” in the original image does not fool a human. Google algorithms are another kettle of fish or booth filled with stuffed ponies.

The Twitter purchase of Madbits (the name suggests swarming or ant methods) delivers some smart folks who have, according to the write up, developed software that:

automatically understands, organizes and extracts relevant information from raw media. Understanding the content of an image, whether or not there are tags associated with that image, is a complex challenge. We developed our technology based on deep learning, an approach to statistical machine learning that involves stacking simple projections to form powerful hierarchical models of a signal.

Once some demonstrations of Twitter’s scaling of this interesting technology is available, I can run the radiation poisoning test. Math is wonderful except when it is not able to do what users expect, hope, or really want to get.

Marketing is good. Perhaps Twitter will allow me to track down this vendor of stuffed ponies. (Yep, it looked real to me.) I know, I know. This stuff works like a champ in the novels penned by Alastair Reynolds. Someday.

Stephen E Arnold, July 30, 2014


Comments are closed.

  • Archives

  • Recent Posts

  • Meta