|
Christopher G. Baker A Block Incremental Algorithm for Computing Dominant Singular Subspaces The Singular Value Decomposition is one of the most useful matrix decompositions, having applications in virtually every scientific and engineering discipline. However, its computation is expensive, prohibiting its use for many large problems. Incremental methods help to lessen this cost by processing the data in groups of vectors, computing an approximation of the dominant singular value decomposition. My research unifies the previous methods incremental methods for
tracking dominant singular subspaces. This general presentation
exposes a new algorithm, with Courses taken outside of major |
a
lower complexity and better runtime performance than previous methods,
over a broader range of algorithm parameters. I also explore the
effect of block size on accuracy of the method. I propose "second
pass" algorithms. These techniques improve the approximation to
the dominant SVD computed by incremental algorithms by making multiple
passes through the data. Finally, I explore the application of the
proposed algorithms to a problem in face recognition and image compression.
