# Reparameterizable Subset Sampling via Continuous Relaxations

## TL;DR

This paper introduces a continuous relaxation method for subset sampling that enables reparameterized gradients, improving training efficiency and performance in various machine learning tasks involving subset selection.

## Contribution

It generalizes the Gumbel-max trick to allow reparameterizable subset sampling, facilitating end-to-end training in tasks like feature selection and neighbor-based models.

## Key findings

- Enhanced feature selection interpretability
- Improved deep stochastic k-NN performance
- Better local neighbor comparison in t-SNE

## Abstract

Many machine learning tasks require sampling a subset of items from a collection based on a parameterized distribution. The Gumbel-softmax trick can be used to sample a single item, and allows for low-variance reparameterized gradients with respect to the parameters of the underlying distribution. However, stochastic optimization involving subset sampling is typically not reparameterizable. To overcome this limitation, we define a continuous relaxation of subset sampling that provides reparameterization gradients by generalizing the Gumbel-max trick. We use this approach to sample subsets of features in an instance-wise feature selection task for model interpretability, subsets of neighbors to implement a deep stochastic k-nearest neighbors model, and sub-sequences of neighbors to implement parametric t-SNE by directly comparing the identities of local neighbors. We improve performance in all these tasks by incorporating subset sampling in end-to-end training.