From 1997 to 2011 he worked as senior computer scientist at NASA Ames Research Center, and became visiting scholar at the Max Planck Institute.
However, in a machine learning context, the theorem makes an implicit artificial assumption regarding the lack of overlap between training and test data that is rarely true in practice.
[14][15] Wolpert has formalized an argument to show that it is in principle impossible for any intellect to know everything about the universe of which it forms a part, in other words disproving "Laplace's demon".
[24] Most prominently, he introduced "stacked generalization",[25] a more sophisticated version of cross-validation that uses held-in / held-out partitions of a data set to combine learning algorithms rather than just choose one of them.
This work was developed further by Breiman, Smyth, Clarke and many others, and in particular the top two winners of 2009 Netflix competition made use of stacked generalization (rebranded as "blending").