Monday, April 07, 2014

Coarse Coding and Coarse Behavior

I confess that I am fascinated with the ongoing saga of Toronto Mayor Rob Ford’s scandalous behavior, his bid for reelection, and the coverage of it all by the Toronto, Canadian, and international media.  I am particularly intrigued when the same event is described completely differently by different news outlets (with competing political aims in their editorial policies).  At the Leafs home game on April 5, was the key story about a happy, lucid, sober Mr. Ford mobbed by a frenzy of adoring supporters?  Or was the news instead about a belligerent, irate, possibly drinking Mayor being warned about his behavior by security?

Regarding this most recent incident, it is of particular interest that social media provided the most detailed account of the Mayor’s evening out.  Posts on Twitter, several including links to video and photographs, tracked his movements from the start of the hockey game to his confrontation with security, alerted followers to his late night visit to City Hall (What was that mysterious burning rubber smell detected by security there? Why is that not mentioned in the traditional media?), and established his early morning presence at the Muzik nightclub.  In this case, it seemed that Twitter was breaking the news, and that the traditional media were playing catchup.  As reports appeared in the media the next day some posts on Twitter accused traditional outlets of not telling the complete story!

All of this made me think of Twitter as representing reality in what connectionist cognitive scientists call a coarse code.  Many artificial neural networks generate highly accurate responses by pooling the signals of individual elements, where each individual element has noisy, sketchy, or inaccurate information about what is going on.  The ‘coarseness’ of this type of representation is reflected in the fact that every processor inside the network is an inaccurate detector.  The surprising power of this representation comes from the fact that if you combine all of these poor measures together, a highly accurate measure is generated.

For coarse coding to work, the individual (inaccurate) measures generally require two different properties.  First, different measuring elements must have overlapping sensitivity: many of them will be measuring similar things.  Second, different measuring elements must also have to have different perspectives on what is being detected.  In short, their sensitivities overlap, but are not identical.  When these two properties are true high accuracy can be produced by combining measures.  This is because if each detector has a different perspective, it will be providing different ‘noise’ than is provided another.  Combining the different noise from different detectors will tend to cancel it all out.  What remains is the amplified ‘signal’ – the ‘truth’ – that is also being sensed to a limited extent by the various processors in the network (due to their overlapping sensitivities).

Each individual tweet on Twitter can be viewed as some information being provided by an inaccurate detector.  If the sources of a large number of these tweets have slightly different perspectives, or provide different kinds of information (statements vs pictures vs videos), then their combined effect provides information that has a strong sense of accuracy.

Not surprisingly, researchers interested in Big Data are actively exploring this characteristic of social media.  For instance, some researchers are using the content of tweets to predict the results of elections, although the accuracy of this approach is subject to a healthy debate.  Importantly, the accuracy of such predictions requires that the two key properties of coarse coding (presenting information that is similar, but different) be true.  When these properties are not true – for instance, when many people retweet the same information, so that different perspectives are not provided – social media can misinform, as shown by Twitter being a continual source of celebrity death hoaxes.

To me, the parallel between tweets and successful coarse coding in artificial neural networks clearly indicates that Twitter can be a source of a great deal of accurate information, and makes me reflect on how neural network paradigms might be tweaked to explore tweet contents.

The parallel also makes me think that if I was a politician seeking reelection – particularly one who is such a notorious celebrity that my frequent encounters with the public immediately appear on social media – I would strive to be on my best behavior.  The image of me emerging from all of those Tweets might be more accurate and telling than the one that the traditional news media feels safe to publish!