Structurally-Constrained Pattern Classifier Design
Signal Compression Laboratory Research Project
Researcher: | Ertem Tuncel |
Faculty: | Prof. Kenneth Rose |
Research Focus: | The design of nearest-prototype (NP)
classifiers is a challenging problem because of the prevalence of poor local minima, and
the piecewise constant nature of the cost function which is incompatible with
gradient-based techniques. In my work, the deterministic annealing (DA) method for
NP-classifier design is extended in two ways. First, the association between prototypes
and class labels is also randomized, and the corresponding association probabilities are
added to the set of parameters to be optimized. Second, the multiplicity (or the mass) of
prototypes are optimized. During the design, all parameters are optimized so as to
minimize the expected misclassification rate for a given level of randomness. The
"joint entropy", which measures the level of randomness, is gradually reduced
while optimizing the cost Lagrangian. As the entropy approaches zero, the method seeks a
deterministic classifier that minimizes the rate of misclassification. The results of the experiments indicate that further significant gains, in terms of better classification performance or reduced classifier complexity, are achieved by randomization and optimization of the mapping between the prototypes and the class labels. |
Presentation: |