O UT - OF -S AMPLE E XTENSION AND R ECONSTRUCTION ON M ANIFOLDS Bhuwan Dhingra Final Year (Dual Degree) Dept of Electrical Engg.
I NTRODUCTION An m- dimensional manifold is a topological space which is locally homeomorphic to the m - dimensional Euclidean space In this work we consider manifolds which are: Differentiable Embedded in a Euclidean space Generated from a set of m latent variables via a smooth function f
I NTRODUCTION n >> m
N ON -L INEAR D IMENSIONALITY R EDUCTION In practice we only have a sampling on the manifold Y is estimated using a Non-Linear Dimensionality Reduction (NLDR) method Examples of NLDR methods –ISOMAP, LLE, KPCA etc. However most non-linear methods only provide the embedding Y and not the mappings f and g
P ROBLEM S TATEMENT x* y* g f
O UTLINE
The tangent plane is estimated from the k- nearest neighbors of p using PCA
O UT - OF -S AMPLE E XTENSION A linear transformation A e is learnt s.t Y = A e Z Embedding for new point y* = A e z* AeAe z*y*
O UT - OF -S AMPLE R ECONSTRUCTION A linear transformation A r is learnt s.t Z = A r Y Projection of reconstruction on tangent plane z* = A r y* z*y* ArAr
P RINCIPAL C OMPONENTS A NALYSIS Covariance matrix of neighborhood: Let be the eigenvector and eigenvalue matrixes of M k Then Denotethen the projection of a point x onto the tangent plane is given by:
L INEAR T RANSFORMATION
F INAL E STIMATES
E RROR A NALYSIS
S AMPLING D ENSITY
N EIGHBORHOOD P ARAMETERIZATION
R ECONSTRUCTION E RROR But A r A e = I, hence
R ECONSTRUCTION E RROR
S MOOTHNESS OF M ANIFOLD
R ESULTS - E XTENSION Out of sample extension on the Swiss-Roll dataset Neighborhood size = 10
R ESULTS - E XTENSION Out of sample extension on the Japanese flag dataset Neighborhood size = 10
R ESULTS - R ECONSTRUCTION Reconstructions of ISOMAP faces dataset (698 images) n = 4096, m = 3 Neighborhood size = 8
R ECONSTRUCTION ERROR V N UMBER OF P OINTS ON M ANIFOLD ISOMAP Faces dataset Number of cross validation sets = 5 Neighborhood size = [6, 7, 8, 9]