Biological and artificial neural networks routinely make reliable distinctions between similar inputs, and the rules for making these distinctions are learned. In some ways, self/nonself discrimination in the immune system is similar, being both reliable and (partly) learned through thymic selection. In contrast to other examples, we show that the distributions of self and nonself peptides are nearly identical but strongly inhomogeneous. Reliable discrimination is possible only because self peptides are a particular finite sample drawn out of this distribution, and the immune system can target the ``spaces'' in between these samples. In conventional learning problems, this would constitute overfitting and lead to disaster. Here, the strong inhomogeneities imply instead that the immune system gains by targeting peptides which are very similar to self, with maximum sensitivity for sequences just one substitution away. This prediction from the structure of the underlying distribution in sequence space agrees, for example, with the observed responses to cancer neoantigens.