tedster - 1:02 pm on Oct 10, 2011 (gmt 0)
The following idea is speculation only. My thoughts have been partially stimulated by the "Panda flux" phenomenon, and also by the idea of machine learning and data modeling for complex goals.
Trying to build a machine model that simulates the human idea of "quality" is a highly adventurous undertaking. We've all seen examples this year of how these early results have missed the mark.
There is another area of human endeavor where machine learning and data modeling is trying to achieve a complex goal - weather forecasting. And that area has several decades of experience built up. If your experience goes back a few decades, you probably realize that short term weather forecasting has improved rather remarkably over that time.
Weather prediction relies on using more than one algorithm - each model takes its data model through to an end result, but then those various predictions get integrated (often by human analysts). You see this especially around predicting hurricane tracks. My point being that a single algorithm is often not enough to accomplish complex goals.
And so I wonder if the "Panda flux" we notice is the result of other supportive algorithms kicking in to balance the main algorithm - perhaps being integrated after some human analysis of the shortfalls and edges cases that surfaced.
The "main Panda algorithm" itself may really be a complex combination of several different models. If that's so, then it would be almost impossible to reverse engineer the thing - and it certainly does seem impossible to take apart right now. In the frustrating absence of some basic understanding, the only guesses we can make are based on very localized and (relatively) small data sets that we each have access to.