Feature Representation and Pattern-filtering for Acoustic Detection

• Activity Number: 11
• Type: Topic Contributed
• Date/Time: Sunday, August 8, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Zhiyi Chi*+
• Companies: University of Chicago
• Keywords: pattern recognition ; point processes ; neuroscience ; bioacoustics ; simulation

Abstract

Many biological studies require detection of various constituents of bioacoustic signals. We developed a fast and accurate approach to the avian acoustic detection. The key is to represent signals by points on the spectro-temporal domain, which significantly reduces the dimension of the data and allows us to formulate acoustic detection as detection of certain global patterns of points. Under a Poisson point process model, the latter detection is formulated as classification based on likelihood ratio and shown to be equivalent to linear filtering of point processes. Accuracy and computational efficiency are achieved by applying multiple filters sequentially, so that only ambiguous ``hot spots'' in the signal are processed in each step. The training of the detector only requires a small sample. It constructs the filters based on structures learned from the sample, and sets the rest of the parameters by simulation. When implemented online, the approach enables neuroscientists to conduct experiments on neurobehavioral interactions with a degree of precision that could not be achieved before.