Topological Acoustical Feature Extraction and Exploitation
Abstract
Although modern deep-learning methods for acoustic target classification can perform well, it is extremely difficult to guarantee th,eir performance. Moreover, the discovery that modern classifiers based on deep learning are excessively sensitive to environmental c,ontext means that models which isolate the target from its context are desperately needed. Traditional filtering and classification,methods rely on stochastic signal models with very specific (usually Gaussian) uncertainty models. While stochasticity is undoubtedl,y necessary, we argue that it isbesides the point when deep learning is used. Deterministic uncertainties about environmental featur,es, sensor trajectory, and target placement will tend to dominate whenever present.The proposed effort will attempt to discover the,fundamental limitations of acoustic scene classification that are classifier agnostic. We propose that successful classifiers are pr,ecisely those that rely on the topological and geometric features of the scene and are additionally insensitive to environmental and, sensor context. Crucially, these features are present regardless of whether the scene is left as raw acoustic data or is processed,into some kind of image. When uncertainties about the collection system are present, we propose that it is better to avoid image for,mation altogether by aggregating all possible environmental contexts into ensemble features.Although our team has led the way in dev,eloping specific topological filters, a fundamental theory of matched topological filtering still remains elusive. Our team is now p,oised to transform our past theoretical success into practical matched filters. We plan to continue to use our successful?but unique,?foundational approach, which applies to all target classification methodologies. The main objective of this effort is to uncover an,d understand classifiable features using the machinery of topological signal processing.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 08, 2022
- Source ID
- N000142212659
Entities
People
- Michael Robinson
Organizations
- American University
- Office of Naval Research
- United States Navy