Jeremy Wyatt Birmingham Why disagreement is a good thing in Neural Net ensembles Committees (or ensembles) of Neural Nets can outperform single networks. This is not surprising. What is surprising is that encouraging the networks in an committee to disagree with one another can improve performance still further. In this talk I'll describe a method called Negative Correlation learning, and explain why it works. Specifically I will show that NC is related to the Ambiguity decomposition introduced by Krogh and Vedelsby. I'll also describe how we can use this insight to find bounds for its parameters. This is joint work with Gavin Brown.