Hey, this is a really good point. However, I don’t think you can do this in Keras because the different loss functions need to be combined into one (usually averaged, see https://keras.io/getting-started/functional-api-guide/#multi-input-and-multi-output-models). This is probably not the case with tensorflow because everything is customisable. I’m happy to be proved wrong though. Also I would give more than just one layer leading upto the final quantile nodes to make them more flexible. Will post this idea as a edit.