Machine learning has gained a lot of interest in the last decade, especially due to impressive advances in deep learning. A typical assumption in machine learning is that the data is i.i.d. from some unknown data distribution. However, in many real-world domains this assumption does not hold, and instead we have some temporal structure in the data. In such cases, it is known that standard optimization algorithms (e.g., SGD) suffer from performance degradation.
Recently, a new optimization technique was proposed for solving optimization problems with Markovian data. In this project, our goal is to implement this algorithm in Pytorch and test its performance when optimizing deep neural networks.