离散与反向传播之间的桥梁：Through-Straight 及其拓展（英文版）

英文标题：Bridging Discrete and Backpropagation: Straight-Through and Beyond中文摘要：本文提出 ReinMax 方法，通过集成 Heun's Method 来解决离散潜变量生成参数的梯度近似问题，该方法在结构化输出预测和无监督生成建模任务中均获得了比现有技术更好的结果，包括 ST 和 ST Gumbel-Softmax。英文摘要：Backpropagation, the cornerstone of deep learning, is limited to computinggradients solely for continuous variables. This limitation hinders variousresearch on problems involving discrete latent variables. To address thisissue, we propose a novel approach for approximating the gradient of parametersi