Sparse distribution of depth soundings in the ocean make it necessary to infer depth in the gaps using alternate information such as satellite-derived gravity and a mapping from gravity to depth. We design and train a neural network on a collection of 50 million depth soundings to predict bathymetry globally using gravity anomalies. We find the best result is achieved by pre-filtering depth and gravity in accordance with isostatic admittance theory described in previous predicted depth studies. When training the model, if the training and testing split is a random partition at the same resolution as the data, the training and testing sets will not be independent, and model misfit results will be too optimistic. We solve this problem by partitioning the training and testing set with geographic bins. Our final predicted depth model improves on old predicted depth model rms by 16%, from 165 m to 138 m. Among constrained grid cells, 80% of the predicted values are within 128 m of the true value. Improvements to this model will continue with additional depth measurements, but higher resolution predictions, being limited by upward continuation of gravity, shouldn’t be attempted with this method.