Discrepancy Directed Model Acquisition for Adaptive Perceptual Systems

For complex perceptual tasks that are characterized by object occlusion and nonstationarity, recognition systems with adaptive signal processing front-ends have been developed. These systems rely on handcrafted symbolic object models, which constitutes a knowledge acquisition bottleneck. We propose an approach to automate object model acquisition that relies on the detection and resolution of signal processing and interpretation discrepancies. The approach is applied to the task of acquiring acoustic-event models for the Sound Understanding Testbed (SUT).