Low-variance Black-box Gradient Estimates for the Plackett-Luce Distribution
This work addresses a bottleneck in optimizing black-box functions over permutations for researchers in machine learning, particularly in causal inference, but it is incremental as it builds on existing variance reduction techniques.
The paper tackled the challenge of high variance in gradient estimates for learning models with discrete latent variables, specifically focusing on latent permutations, and proposed control variates for the Plackett-Luce distribution to enable stochastic gradient descent optimization, showing that it outperforms competitive relaxation-based methods on causal structure learning tasks.
Learning models with discrete latent variables using stochastic gradient descent remains a challenge due to the high variance of gradient estimates. Modern variance reduction techniques mostly consider categorical distributions and have limited applicability when the number of possible outcomes becomes large. In this work, we consider models with latent permutations and propose control variates for the Plackett-Luce distribution. In particular, the control variates allow us to optimize black-box functions over permutations using stochastic gradient descent. To illustrate the approach, we consider a variety of causal structure learning tasks for continuous and discrete data. We show that our method outperforms competitive relaxation-based optimization methods and is also applicable to non-differentiable score functions.