What if the gradient is biased?


#1

Is there a pytorch learner that will work if the gradient of the ELBO with respect to φ is a uniformly biased estimate? That is, if the difference between the gradient estimates at different points is an unbiased estimate of the difference between the true gradients at those points, but the only way to estimate the first-order bias is numerically? Being able to use such a learner instead of Adam would make a lot of the rest of what I’m doing a lot easier and faster…