Interactive demo · Probabilistic Machine Learning · Chapter 20

Isomap on a Swiss roll: neighbours and short circuits

Isomap connects each point to its \(K\) nearest neighbours, measures distance along the graph (shortest paths), and applies classical MDS to those distances. On the left the roll is seen from the side, with the graph drawn: red edges jump between layers (their positions along the roll differ by more than 3 radians) and are the short circuits. A few red edges are enough to ruin the unrolled picture on the right; too few neighbours (K = 2 here) split the graph into pieces. Classical MDS on the plain Euclidean distances is PCA and cannot unroll the roll at all.

Method

Neighbours K

Noise (standard deviation added to each coordinate)

Try these

The roll has 500 points (the scikit-learn construction, positions \(t\in[1.5\pi,4.5\pi]\), width 21). The unrolling score is the absolute correlation of the first embedding coordinate with the position \(t\) along the roll. The top eigenvectors are found by subspace iteration, so the score is reliable when the top two eigenvalues are well separated from the third.

In the lecture notes