Edited September
17, 2004
Anticipatory Web
5/6/2004
2:22 PM
Communication from Paul Werbos
Good afternoon!
I would interpret Paul Prueitt’s response to my question as follows.
There are really three kinds of data bases out there, roughly.
There are time-series databases, like many speech or econometric time-series, from which it is valid to do training or inference to develop dynamic models, which in turn are valid for prediction.
There are truly cross-sectional databases, "snapshots of the world" -- from which one can compute correlations or other static properties analogous to correlations -- but it is GROSSLY incorrect to make predictions of the future based solely on what one learns from such patterns. (In first year macroeconomics, for example, they show how cross-correlations seem to imply an elasticity of consumption with respect to income of 0.7, when in time series it is 1.0.In practical terms, this means you will make a 30 percent error in predicting changes in consumption if you based it on static analysis. Similar errors occur in large "sophisticated" energy forecasting models. Worse errors occur in many Policy-oriented models.)
But there is something intermediate, where a database "tries" to develop a distilled version of past time series, so as to make inference still possible but to reduce the storage requirements compared with true time-series.
When one can afford true time-series, they are far better.
But it is clear that the brain does NOT use time-series.
Yet it is also clear that the brain can "learn from memory." We can relive past experience, reinterpret it, and adapt accordingly. The imperfections of our memory reduce the benefits of this kind of adaptation, but still it is far better than having no memory at all. This is a case where I would even refer back as far as Freud, and Pribram's book on Freud.(Or even to Barrett and Yankelovich''s easy paperback, which, though an exercise in yellow journalism, does convey the key idea here.)
This kind of principle is important for neural networks and machine learning, of course, and is not a small research issue.
===========
But I was asking the whole group... about.. a more worrisome situation...in which people try to make predictions (e.g. "threat anticipation") based on the second type of database, which is something of a disaster...if we lose our memory, it is hard...
=======================
Best of luck,
Paul Werbos