Gradient boosting is a methodology that can be applied on another machine learning algorithm (think MSE). It typically involves two types of models: 1. A weak machine learning model. 2. A strong model built out of many weak models. The process involves choosing a new weak model at each iteration of the process and using it to predict the error of the current strong model.This is referred to as pseudo response. According to Google, "Error is the difference between the prediction and a regressive label". This operation will repeat until a stopping condition is met (e.g. a max number of iterations or the model begins to overfit).