METHODS FOR ANOMALY DETECTION

The majority of machine learning algorithms train their models by solving optimization problems. Typical objective functions are non-convex and NP-hard to optimize. The complexity of the optimization problem is exacerbated by operating in big dimensions, and it is often essential that computationally inexpensive algorithms are used. This project will investigate algorithms and methods for optimization problems which arise in machine learning. The most popular algorithm is gradient decent, but there are problems with estimating the gradient. Amongst others, this project will consider data-driven estimates of the gradient.