An introduction to unsupervised machine learning. In this presentation, I go over the different data representations for unsupervised learning before covering clustering, dimensionality reduction (including t-SNE and UMAP), and various types of autoencoders. Finally, I discuss the difficulties in evaluating results from unsupervised learning and the similarities between seemingly different unsupervised learning methods. If you are not already familiar with the basics of machine learning, I recommend that you first watch my short introduction to the core concepts: • Machine learning: A sh...
0:00 Introduction: definition of unsupervised learning and overview of the presentation
0:38 Input data: high-dimensional data, distance matrix, similarity matrix, and network representations
1:21 Clustering: hierarchical clustering, partitional clustering, k-means clustering, and MCL
2:15 Dimensionality reduction: latent representation, principal component analysis (PCA), t-SNE, UMAP, multidimensional scaling (MDS), force-directed network layouts, and their similarities
3:17 Autoencoders: predicting input from input, neural networks with bottleneck, linear autoencoders, denoising autoencoders, and variational autoencoders
4:54 Evaluation: no ground truth, consistency of labels, intra-cluster cohesion, inter-cluster separation, percent variance captured, independent test sets for autoencoders, and downstream benchmarking
6:38 The big picture: hierarchical vs. partitional clustering, linear autoencoders vs. PCA, t-SNE/UMAP cluster visualization, and network layout vs. dimensionality reduction