You can't, not yet. However you can specify the following: >The main purpose of the system is to take a program describing a stochastic computation graph and t

You can't, not yet. However you can specify the following: >The main purpose of the system is to take a program describing a stochastic computation graph and to run a randomized algorithm (stochastic backpropagation) that, in expectation, provably generates unbiased samples of the gradients of the loss function with respect to the parameters. which is pretty good, and can save you from errors in manipulating the math of your gradient to make it have less variance, and in implementing it.