Implementation of Active Learning Methods on Poisson Learning Framework
The objective of this UROP project is to implement three active learning methods (V-Optimality, Error Bounds, and Graph Signals) to Poisson Learning and Laplace Learning to see how they improve label selection and accuracy at low label rates. The approach we took for this topic is by first implementing the active learning methods to a toy dataset so that we can visualize how the method is performing. Once the implementation of the methods is satisfactory, we implement them into real-world datasets (MNIST or MSTAR). From the three methods proposed, we implemented V-Optimality and Graph Signals successfully but the Error Bounds method had mathematical derivation issues in the paper which led us to discard implementing the method. The conclusion we got out of this project is that active learning certainly helps improve graph-based semi-supervised learning accuracy at low label rates. However, the success of an algorithm depends on the dataset we use for our tests.