Efficient Learning of Discrete-Continuous Computation Graphs

TL;DR

This paper analyzes complex stochastic computation graphs with multiple discrete components, identifies training challenges, and proposes strategies like Gumbel noise scaling and dropout residuals to enable effective training and better generalization.

Contribution

It introduces novel training strategies for complex discrete-continuous models, overcoming gradient issues and enabling training of more intricate stochastic computation graphs.

Findings

Proposed Gumbel noise scaling improves training stability.

Dropout residual connections enhance optimization of discrete-continuous models.

Complex models outperform simpler counterparts on benchmark datasets.

Abstract

Numerous models for supervised and reinforcement learning benefit from combinations of discrete and continuous model components. End-to-end learnable discrete-continuous models are compositional, tend to generalize better, and are more interpretable. A popular approach to building discrete-continuous computation graphs is that of integrating discrete probability distributions into neural networks using stochastic softmax tricks. Prior work has mainly focused on computation graphs with a single discrete component on each of the graph's execution paths. We analyze the behavior of more complex stochastic computations graphs with multiple sequential discrete components. We show that it is challenging to optimize the parameters of these models, mainly due to small gradients and local minima. We then propose two new strategies to overcome these challenges. First, we show that increasing the scale parameter of the Gumbel noise perturbations during training improves the learning behavior. Second, we propose dropout residual connections specifically tailored to stochastic, discrete-continuous computation graphs. With an extensive set of experiments, we show that we can train complex discrete-continuous models which one cannot train with standard stochastic softmax tricks. We also show that complex discrete-stochastic models generalize better than their continuous counterparts on several benchmark datasets.