Dimensionality reduction using Haar wavelets Retain sum and lower level detail coefficients For example, retain 1 sum (at level 0) and 3 detail (1 at level 0 and 2 at level 1) coefficients Contractive Alternatively, for a database of objects, retain those coefficients whose variances are highest Another option is to zero out coefficients whose absolute values are below a threshold What happens when dimensionality is not a power of 2?   Pad zeros at end: latter half becomes less important   Pad equal amount of zeros in each half: should be done recursively Interestingly, shuffing the dimensions produce different wavelet coefficients   How to shuffle to satisfy some critera?