Abstract
The goal of most learning processes is to bring a machine into a set of "correct" states. In practice, however, it may be difficult to show that the process enters this target set. We present a condition that ensures that the process visits the target set infinitely often almost surely. This condition is easy to verify and is true for many well-known learning rules. To demonstrate the utility of this method, we apply it to four types of learning processes: the perceptron, learning rules governed by continuous energy functions, the Kohonen rule, and the committee machine.
Original language | English |
---|---|
Pages (from-to) | 1839-1861 |
Number of pages | 23 |
Journal | Neural Computation |
Volume | 13 |
Issue number | 8 |
DOIs | |
State | Published - Aug 2001 |