Analysis of gradient-based optimization objective for robust machine learning classification

The idea behind creating artificial intelligence extends far back in human history, founded on the idea of imitating human learning to better predict and make decisions. For over 70 years, scientists have worked on developing the rational ability in computers but are still a long way from the ultimate goal of creating a machine able to outperform humans in every intelligence-limited task. This thesis builds on top of state-of-the-art research in finding robust models to predict out-of-distribution data better. Most research has been done in generalizing for linear regression and my hypothesis is that the same ideas can be extended to classification using logistic regression. This thesis builds on a gradient-based-risk optimization objective named CoCo. The hypothesis has been tested by visualizing traditional machine learning loss functions versus the gradient-based risk for different causal structures. The results show that similar outcomes can be seen in linear regression compared to logistic regression, implying that it may be viable in classification.