| Despite the variety of statistical classifiers available, many inductive classifiers are ultimately based on exponential functional forms. These models are quite versatile and perform well in a variety of situations, but they also have a strong tendency towards overconfidence. They will typically be quite sure of their predictions, whether correct or not, except when quite close to a decision boundary.;This dissertation, motivated by the Vocal Joystick project, an assistive device allowing individuals with motor impairments to operate mouse pointers or other electromechanical devices using non-verbal vocalizations, examines a family of statistical classification models designed to allow smoother transitions between classes.;We first present a classification model formulated as a ratio of semi-definite polynomials, which we call the ratio semi-definite classifier (RSC). The RSC allows smoother transitions across class boundaries, and does not demonstrate the overconfidence bias typical of models based on ratios of exponentials Testing on several corpora of various sizes, we find that the RSC performs well, but often has slightly lower accuracy than an exponential model such as a multi-layer perceptron (MLP).;To improve the accuracy of the RSC while retaining its other properties, we propose several extensions to the model, creating a family of RSC-based classifiers. We test two other members of that family, the multi-layer RSC (ML-RSC) which adds a hidden layer to the model allowing us to learn a kernel-like feature transformation in a data-driven manner. Experimental results show that the ML-RSC provides superior accuracy compared with the original RSC and an MLP.;We also propose a semi-supervised learning (SSL) framework suitable for any differentiable, parametric and optionally multiclass model. With this framework, we demonstrate improved results with semi-supervised versions of the RSC, ML-RSC and MLP on several data sets.;Finally, we present a comprehensive look at the Vocal Joystick engine, describing its architecture and the various signal processing and machine learning techniques used in the system. This includes a discussion of where the RSC and ML-RSC, and adaptation of the parameters for those models, are used in the Vocal Joystick. |
