Synopsis

While medical problems are typically associated with the binary classification task (i.e., disease versus healthy controls), Glaucoma is characterized by a continuum of conditions affecting the optic nerve. Central to this research are the analysis of shape information and visualization of disease development, with particular interest in rates of deterioration (e.g., nerve damage and correspondence with disease trajectory). To this end, shape-based appearance models of optic nerve tomography are employed in the design of a data-driven system that provides visual representations of optic nerve changes and relationships with disease subtype.

Synopsis

This research concerns the development of specialized Functional MRI (fMRI) data manipulation and analysis tools and investigates procedures appropriate for detecting and visualizing early indications of change in functional connectivity in the context of Alzheimer's Disease. More specifially we are interested in building effecient algorithms for automatic and objective detection of changes in the brain's resting state networks and neurocompensatory response during disesase progression.

Novelty Detection + Coevolution = Automatic Problem Decomposition:
A Framework for Scalable Genetic Programming Classifiers [ pdf ]

Abstract

A novel approach to the classification of large and unbalanced multi-class data sets is presented where the widely acknowledged issues of scalability, solution transparency, and problem decomposition are addressed simultaneously within the context of the Genetic Programming (GP) paradigm. A cooperative coevolutionary training environment that employs multi-objective evaluation provides the basis for problem decomposition and reduced solution complexity, while scalability is achieved through a Pareto competitive coevolutionary framework, allowing the system to be readily applied to large data sets (tens or hundreds of thousands of exemplars) without recourse to hardware-specific speedups. A key departure from the canonical GP approach to classification involves expressing the output of GP in terms of a non-binary, local membership function (e.g., a Gaussian). Decomposition is achieved by reformulating the classification problem as one of cluster consistency, where an appropriate subset of the training patterns can be associated with each individual such that problems are solved by several specialist classifiers as opposed to a singular `super' individual.

Although multi-objective methods have previously been reported for GP classification domains, we explicitly formulate the objectives for cooperative behavior. Without this the user is left to choose a single individual as the overall solution from a front of solutions. This work is able to utilize the entire front of solutions without recourse to heuristics for selecting one individual over another or duplicating behaviors between different classifiers.

Extensive benchmarking was performed against alternative frameworks for classification including Genetic Programming, Neural Networks, and deterministic methods. In contrast to classifiers evolved using competitive coevolution alone, we demonstrate the ability of the proposed coevolutionary model to provide a non-overlapping decomposition or association between learners and exemplars, while returning statistically significant improvements in classifier performance. In the case of the Neural Network methods, benchmarking is conducted against the more challenging second order neural learning algorithm of conjugate gradient optimization (previous comparisons limit Neural Networks to first order methods). The proposed evolutionary method was often significantly better than the non-linear Neural Network, whereas the linear model tended to work well or not at all. In effect, the evolutionary paradigm provided a more robust model for searching the space of non-linear models than provided under the neural gradient decent paradigm. With respect to deterministic methods, the problem of benchmarking stochastic versus deterministic algorithms is first addressed, with a new methodology established for making such comparisons. The ensuing comparison demonstrated that the evolutionary algorithms remain competitive with most data sets appearing to benefit from the proposed evolutionary methodology.

Synopsis

In this work we investigate several issues of relevance to data driven learning in the context of intrusion detection systems (IDS) including: scaling to large data sets, problem decomposition, class imbalance, and solution transparency. Competitive coevolution provides the basis for scaling evolutionary methods to arbitrarily large and unbalanced data sets, while the multi-objective classifier framework provides the ability to decompose a problem into a series of smaller simpler solutions and establishes the basis for highly modular solutions that are applicable to real time IDS operations.

Synopsis

In this research we applied a novel classification approach to categorize digital optic nerve images into one of several pre-defined classes, indicating healthy or various subtypes of optic nerve damage patterns. We were also interested in establishing the degree of confidence associated with each recommendation and the underlying rule from which classifications are derived. To this end we developed a classifier based on Multi-objective Genetic Programming, which provided the basis for establishing analytic solutions and automatic problem decomposition.

Synopsis

This project was mainly concerned with automatic feature extraction and indexing of high resolution optic nerve images that were gathered in a clinical environment using CSLT imaging technology. The combination of objective imaging, shape-based feature extraction and classification techniques examined in this research allowed automatic and objective discrimination between healthy and diseased nerves. Our scheme compared favorably to classification using conventional stereometric feature sets and completely removes the highly subjective traditional requirement for manually identifying the edge of the optic nerve. Moreover, the results of this study enabled us to reduce the 256 shape-based features to 6.