Abstract by Jared Neilson
Semi-Supervised Learning with UMAP Embeddings
In machine learning, obtaining sufficient labeled data can be expensive or infeasible. Semi-supervised learning (SSL) tackles this problem by using the distribution of unlabeled data to augment limited labeled data. We apply the unsupervised UMAP algorithm to create fast, descriptive embeddings for data preprocessing. This approach yields improved test accuracy and faster training time on state-of-the-art SSL algorithms. We demonstrate improvements across several synthetic and realistic datasets.