For Monday Read chapter 12 Program 4 Any questions? Visualizing - PowerPoint PPT Presentation

For Monday • Read chapter 12

Program 4 • Any questions?

Visualizing Weight Vectors • 2d network topology

• 2-D input space

Self-organization of weight vectors • weight vector is a point on the unit square • connected to the four nearest neighbors • Process: • 0 samples: random • 30 samples: groups formed • 100 samples: groups organized • 10,000 samples: individual units laid out • 2D->2D SOM demo

Sliders • http://www.cis.hut.fi/research/javasomdemo/ demo1.html

Colors Demo • http://www.superstable.net/sketches/som/

Useful? • High dimensional inputs (similarities hard to see) • maps to 2-D classification (similarities evident) • The 2 output dimensions abstracted automatically to provide best discrimination

Example • Map of phonemes • Input: energy in 10 different frequency ranges • Output: a 2-D map of the phonemes of the language

Approximating 2D in 1D

Approximating 3D in 2D • 3D->2D SOM demo

Learning Issues • Alpha • Neighborhood size

Genetic Algorithms • Inspired by evolution • Really a form of search

Basic Concept • Have a pool of current solutions • Mate those solutions to get new solutions • Mutate some solutions • Remove the worst solutions

Chromosomes • Representation of a solution to the problem • Generally in the form of a bit-string • Determining the encoding of a solution into a string is an important step

Crossover • The mating between two chromosomes, generally producing two new chromosomes

Mutation • Simply random modification of chromosomes • Generally at much higher rates than natural mutation (1-3%, typically)

Fitness Function • Measure of the quality of the solution • Determines how likely the solution is to survive to the next generation • Determines how like the solution is to be selected for crossover