They actually use a relu to represent the model weights. But I'm not convinced that this can't be avoided. We do gradient boosted decision tree training without

They actually use a relu to represent the model weights. But I'm not convinced that this can't be avoided. We do gradient boosted decision tree training without this trick.