Re-evaluation of the Gottfried sum using neural networks

We provide a determination of the Gottfried sum from all available data, based on a neural network parametrization of the nonsinglet structure function F_2. We find S_G=0.244 +- 0.045, closer to the quark model expectation S_G=1/3 than previous results. We show that the uncertainty from the small x region is somewhat underestimated in previous determinations.