Seminar Details
2010-09-21 (15.00) : Optimization algorithms tailored for signal/image processing applications
At Euler building
Organized by Mathematical Engineering
Speaker :
Suviseshamuthu, Easter Selvan
Abstract :
The talk concerns a few applications of biologically-inspired/heuristic optimizers as well as a collection of deterministic optimization
algorithms on a manifold setting. First, it is demonstrated how a breast density model can be built, which serves as a breast cancer risk
assessment tool, by estimating the parameters of breast tissues'density distributions with the aid of heuristic optimizers; the segmentation
accuracy is shown to surpass the results obtained using the expectation maximization (EM) algorithm. Next, incorporating a trivial technique
for implicit imposition of the orthonormality constraint while optimizing an independence criterion with well-known global optimizers is
concisely presented; this application illustrates a way to handle a constraint in a more natural way with unconstrained optimizers suitable
for non-convex landscapes. Finally, a collection of deterministic optimizers on a trivial manifold is shown to improve the estimation accuracy of
unmixing matrix, as a consequence of relaxing the orthonormality constraint in the blind source separation problem.