sheng yu um statistics. outline motivation strategy sample algorithms

22
A Strategy for Manifold Clustering with Sample Algorithms Sheng Yu UM Statistics

Upload: lane-upham

Post on 31-Mar-2015

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

A Strategy for Manifold Clustering with Sample Algorithms

Sheng YuUM Statistics

Page 2: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Outline

Motivation

Strategy

Sample Algorithms

Page 3: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Motivation (pattern)

Most current clustering methods are only able to detect agglomerated patterns.

New generation methods, such as normalized cut, have more flexibility, but are still not able to detect twisted, perhaps also entangled manifolds.

Such manifold patterns are not rare.

Page 4: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

This is a manifold...

Page 5: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Example:

Try to cluster a pair of symmetric double spiral.

Page 6: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Example:

result fromk-means

Page 7: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Example:

result from normalized cut

Page 8: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Motivation (noise)

Theoretically, hierarchical clustering method using “single linkage” as the merging criterion is able to cluster twisted patterns. However, since “single linkage” is extremely sensitive to noisy, it is not actually a usable method.

Page 9: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

MotivationTo design a new method that is not only able

to accomplish traditional “easy” tasks, but also handles twisted, entangled patterns as well.

Also, this new method should not be ruined by noise (moderate level, in terms of signal-noise ratio).

Page 10: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Outline

Motivation

Strategy

Sample Algorithms

Page 11: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Strategy (rationale)

Page 12: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Strategy (design)Engine: Searches

paths between each pair of points. More powerful engine provides faster speed.

Filter: Tells the engine which neighboring points can be connected from a specific start point. Controls the quality.

Engine

Filter

Page 13: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Example (easy one)

Page 14: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Example (not so easy one)

Page 15: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Example (not so easy one)

Page 16: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Example (hard one)

Page 17: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Outline

Motivation

Strategy

Sample Algorithms

Page 18: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Algorithms (filter)The filter I currently use is still primitive.

But it does a lot of jobs, such as the above examples.

The strategy is an open framework. We can build better filters to detect even more difficult patterns and have more resistance to noise.

Page 19: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Algorithmsthe importance of the engine

Sample Size 320

Sample Size 640

Brute force 97” Death touch

Fission 1.5” 65’’

Algebraic fission

Never minded 0.5”

Page 20: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

Strategy (rationale)

Page 21: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

AlgorithmsThe true benefit of a super fast engine is that

it allows us to do iteration.

We need to set up a range of acceptable number of clusters.

We do not need our initial parameters to be precise. The algorithm will do heuristic search for us.

Page 22: Sheng Yu UM Statistics. Outline Motivation Strategy Sample Algorithms

a demo of visual aids for choosing parameters