3 ms·
The convergence rate on learning from "x didn't work" is much slower than from "x did work". If you're looking for a peak, most of the time you want to climb t
by yodon 8y ago
The convergence rate on learning from "x didn't work" is much slower than from "x did work".
If you're looking for a peak, most of the time you want to climb the gradient rather than descend it. That statement is true even if you suspect there are multiple peaks in the landscape.
- julvo 8y agoLearning from negative examples can be more effective when we can conclude that all inputs didn't work. The problem with positive examples is that we don't know which of the inputs was responsible for the positive result.