Improved lower bounds for learning from noisy examples: an information-theoretic approach