Optimizing with constraints: reparametrization and geometry.

When training machine learning models, and deep networks in particular, we typically use gradient-based methods. But if we require the weights to satisfy some constraints, things quickly get more complicated. Some of the most popular strategies for handling constraints, while seemingly very different at first sight, are deeply connected. In this post, we will explore these connections and demonstr

[misshiki]

制約付きの最適化について。勾配降下ではなく鏡像降下という方法など。コードはPyTorch。