We analyze the ''query by committee'' algorithm, a method for filterin
g informative queries from a random stream of inputs. We show that if
the two-member committee algorithm achieves information gain with posi
tive lower bound, then the prediction error decreases exponentially wi
th the number of queries. We show that, in particular, this exponentia
l decrease holds for query learning of perceptrons.