GO Gradient for Expectation-Based Objectives
This addresses a fundamental bottleneck in gradient estimation for various ML algorithms, offering a general solution that could impact methods relying on expectation-based objectives, though it appears incremental by building on existing gradient techniques.
The paper tackles the problem of efficiently computing unbiased gradients for expectation-based objectives in machine learning, proposing a General and One-sample (GO) gradient that applies to many distributions, including non-reparameterizable ones, and achieves low variance comparable to the reparameterization trick, often working well with just one Monte Carlo sample.
Within many machine learning algorithms, a fundamental problem concerns efficient calculation of an unbiased gradient wrt parameters $\gammav$ for expectation-based objectives $\Ebb_{q_{\gammav} (\yv)} [f(\yv)]$. Most existing methods either (i) suffer from high variance, seeking help from (often) complicated variance-reduction techniques; or (ii) they only apply to reparameterizable continuous random variables and employ a reparameterization trick. To address these limitations, we propose a General and One-sample (GO) gradient that (i) applies to many distributions associated with non-reparameterizable continuous or discrete random variables, and (ii) has the same low-variance as the reparameterization trick. We find that the GO gradient often works well in practice based on only one Monte Carlo sample (although one can of course use more samples if desired). Alongside the GO gradient, we develop a means of propagating the chain rule through distributions, yielding statistical back-propagation, coupling neural networks to common random variables.