Non-linear dimension reduction techniques have more degrees of freedom than linear techniques. That enables them to better preserve small distances, which is typically more important than preserving middle and large distances. Essential examples are multidimensional scaling (MDS) and stochastic techniques, such as Stochastic Neighborhood Embedding (SNE) and its variant t-SNE. These are discussed along with their parameters. We have a look at examples where the stress value is computed – an essential quality metric.
Chapters:
00:00 - Outline
02:09 - Non-Linear Projection Techniques
25:43 - Stochastic Neighborhood Embedding
36:11 - t-SNE
47:45 - Crafted Projections
51:40 - Assessment of Projection Quality
