Pattern Filtering for Acoustic Recognition Using Point Processes

• Activity Number: 59
• Type: Invited
• Date/Time: Sunday, August 8, 2004 : 6:00 PM to 7:50 PM
• Sponsor: IMS
• Author(s): Zhiyi Chi*+
• Companies: University of Chicago
• Keywords: pattern recognition ; point processes ; neuroscience ; bioacoustics ; simulation

Abstract

Many scientific studies require fast and accurate pattern recognition. In the case of acoustic detection, this is challenging because of the variations in acoustic patterns. To address this, an important idea is to construct representations of the inputs that are robust to variations. We find that the structures of bio-acoustic signals can be represented by points on the frequency-time domain. The resulting, discrete, representations not only exhibit robustness to random variations, but also significantly reduce the dimension of the data. We then can formulate acoustic detection as detection of global patterns of points. Under a Poisson point process model, the latter can be achieved by linear filtering of point processes. The training of the detector only requires a small sample. It constructs filters based on the structures learned from the sample, and tunes the other parameters by simulation. We have implemented the approach both off-line and on-line. It has enabled neuroscientists to conduct experiments on neuro-behavioral interactions with a degree of precision that could not be achieved before.