Active Learning as Non-Convex Optimization
Andrew Guillory, Erick Chastain, Jeff Bilmes; JMLR W&CP 5:201-208, 2009.
We propose a new view of active learning algorithms as optimization. We show that many online active learning algorithms can be viewed as stochastic gradient descent on non-convex objective functions. Variations of some of these algorithms and objective functions have been previously proposed without noting this connection. We also point out a connection between the standard min-margin offline active learning algorithm and non-convex losses. Finally, we discuss and show empirically how viewing active learning as non-convex loss minimization helps explain two previously observed phenomena: certain active learning algorithms achieve better generalization error than passive learning algorithms on certain data sets and on other data sets many active learning algorithms are prone to local minima.