next up previous contents
Next: Temporal mismatch Up: How it is different Previous: Many featuresnot enough

Strong attribute correlation

In many learning systems, we assume that the attributes are independent in some way; if not independent, then at least measuring different observations of the object we are trying to recognise. In this case, however, we know that the attributes are not independent; in fact there is usually a high correlation. For example, consider a position sensor on a robot. We know that due to physical constraints, the position at time t+1 is correlated with the position at time t.

In fact, we can take advantage of the fact by making our temporal learning algorithms use the correlations to reduce the data significantly.



Mohammed Waleed Kadous
Tue Oct 6 13:04:40 EST 1998