CSC 2535: Lecture 10

Non-parametric, non-linear dimensionality reduction

Dimensionality reduction: Some Assumptions

The basic idea of non-parameteric dimensionality reduction

Two types of dimensionality reduction

Linear methods of reducing dimensionality

Metric Multi-Dimensional Scaling

Converting metric MDS to PCA

Other non-linear methods of reducing dimensionality

Problems with Sammon mapping

IsoMap: Local MDS without local optima

How Isomap measures intrinsic distances

Using Isomap to discover the intrinsic manifold in a set of face images

Slide 13

Maps that preserve local geometry

Finding the optimal weights

A sensible but inefficient way to use the local weights

Local Linear Embedding: A less sensible but more efficient way to use local weights

The convex optimization

The collapse problem

Failure modes of LLE

A comment on LLE

Maximum Variance Unfolding

Stochastic Neighbor Embedding

A probabilistic local method

Throwing away the raw data

Evaluating an arrangement of the data in a low-dimensional space

The cost function for a low-dimensional representation

The forces acting on the low-dimensional points

Data from sne paper

Picking the radius of the gaussian that is used to compute the p’s

Symmetric SNE

Computing the p’s for symmetric SNE

Turning conditional probabilities into pairwise probabilities

Evaluating an arrangement of the points in the low-dimensional space

The cost function for a low-dimensional representation

The forces acting on the low-dimensional points

Slide 37

Optimization methods for SNE

More optimization tricks for SNE

A more interesting variation that uses the probabilistic foundation of SNE

A nice dataset for testing “Aspect maps”

Slide 42

The relationship between aspect maps and clustering

A weird behaviour of aspect maps

Why SNE does not have gaps between classes

Slide 46

Slide 47

Slide 48

Slide 49

Slide 50

Slide 51

t-SNE

Optimizing t-SNE