could you please introduce some recent papers on this? How to deal with the (approximate) gradient descent for graphical model inference, and then integrate it into the stochastic gradient descent? I know the Domke's workhttp://machinelearning.wustl.edu/mlpapers/papers/AISTATS2012_Domke12 was devoted on this. And the structured prediction energy networks is also on this https://arxiv.org/abs/1703.05667. But the mathematical requirements for both two papers is beyond mine extreme. Could you please show some simpler methods with simpler math derivations? Thanks a lot!