Thursday, June 26, 2008

The End of Theory

Thanks to a post on the KurzweilAI blog, I read this Wired story:

The highly thought-provoking article describes some uses of high-volume data mining, and makes some extremely valuable points. It uses the obvious success of Google as a case in point: they exploit the links in the WWW as a value measure, and consequently provide a greatly improved search engine compared to the old ways. Note that Google does not seek or care to evaluate the links themselves, the pages they point to, or what they mean, only that the existence of links and their patterns at a large enough scale reflect the utility of pages.

This concept can be expanded in a myriad ways. For example, put cameras in stores (grocery, shoe, clothing, it doesn't matter). Note the correlations between shoppers pausing at a shelf or display, and putting items in a cart. Given enough data, you don't care about the identity of the shopper or whether they actually purchased the items or not. You can still infer information such as sizes, styles, and tastes from their shopping patterns, and then modify advertisements (and/or sale coupons) based upon what they are likely to buy or what competing products may be selected.

Voice recognition and language translation are already succumbing to similar attacks: you don't need to understand how speech works (or how a language is structured) to solve those problems: you simply need massive amounts of applicable data and a large (and fast) enough neural network.

Will similar techniques (given cameras in public places including stores and banks) allow a thief to be identified before he/she acts simply based upon their behavior? I think so. I also hope they don't mistakenly think I might be a thief.

I'll bet that someone is already making millions in the stock and commodities market by analyzing billions of trades, without bothering or needing to understand who bought what. It's only the pattern that counts.

What are the implications for behavioral influence? Can people be effectively controlled by feeding us the inputs that statistically result in targeted behaviors? And not just on average, but by relating to the individuals past response patterns?

And you were worried about the loss of privacy!

No comments: