Dissertation Spatial Hearing, Auditory Sensitivity, and Pattern Recognition in Noisy Environments

This dissertation investigates how human perceivers and learning algorithms cope with noisy, complex environments. The first part of the dissertation presents results of psychoacoustic experiments that give insight into human auditory perception of sounds originating near the listener. The second part develops a memory-based computational learning architecture that is resistant to noise while limiting its own internal complexity. The first part of the dissertation describes results of two psychoacoustic experiments. The first experiment investigates the ability of human listeners to employ binaural auditory processing when detecting a sound from various positions around the listener. The experiment is performed in a simulated noisy anechoic environment, which provides more realistic spatial cues than have been tested in previous headphone studies of masked sound detection while allowing the acoustic signals to be fully characterized. The results support the idea that performance is determined by brainstem auditory processing and is not influenced by any higher-level sound-location-based processing. The second psychoacoustic experiment studies how the listener’s position in a room and the listener’s experience in a room influence the ability to accurately localize nearby sounds in azimuth and distance. The results show that the listener position in the room influences response variability and that as a listener becomes familiar with a room, passive learning occurs, reducing response variability. Finally, the presence of a nearby wall has a small influence on azimuthal response bias, but no noticeable effect on perceived source distance. In the second part of the dissertation, a memory-based learning system, called PointMap, is proposed. This system implements new methods of pruning based on informative value of the stored memories. It is shown that such a “forgetting” mechanism can solve the problems of category proliferation and noise sensitivity in many learning systems. Variations on the Adaptive Resonance Theory systems are also investigated. Both portions of this dissertation demonstrate that learning enables humans and machines to perform robust pattern recognition tasks in the face of noise and uncertainty in the inputs.