These guys proposed droupout in 1992. The differences are that Clay and Séquin dropped only units not inputs, and that they did not normalize the learned weights after completing training.
"Fault Tolerance Training Improves Generalization and Robustness"
Reed D. Clay and Carlo H. Séquin
[Proceedings 1992] IJCNN International Joint Conference on Neural ...
The @the Europeist comment is correct - Google is showing references from patent publications. Technically they are not really citations until they appear printed in a patent. It looks like Google is going into PAIR or the Global Dossier to the file wrapper and getting information from the IDS forms filed in the applications.