Processing Acoustic Scenes with Topology (PAST)

Abstract

Although modern deep-learning methods for acoustic target classification can perform well, it isextremely difficult to guarantee their performance. To remedy this deficiency, the proposedeffort will attempt to discover the fundamental limitations of acoustic scene classification thatare classifier agnostic. We propose that successful classifiers are precisely those that rely on thetopological and geometric features of the scene. Crucially, these features are present regardlessof whether the scene is left as raw acoustic data or is processed into some kind of image. Whenuncertainties about the collection system are present, we propose that it is better to avoid imageformation altogether. For instance, although many other researchers have examined sonar targetclassification in synthetic aperature sonar, our approach is centered around low-dimensionalmodels derived from ""raw"" (high-dimensional) sonar echos.Traditional filtering and classification methods rely on stochastic signal models with veryspecific (usually Gaussian) uncertainty models. While stochasticity is undoubtably necessary, weargue that deterministic uncertainties about sensor and target placement in the scene will tend todominate if present. To properly address these deterministic uncertainties both in the scene andin the collection system, our project will leverage three related mathematical subdisciplines, noneof which are typically employed in sonar data processing:1. Manifolds and sheaves to represent target signatures,2. Algebraic topology to extract features from target signatures, and3. Topological filters to describe signal processing chains.Topological methods are particularly well-aligned with the goal of gaining insight into physicalprocesses, since they highlight symmetries which are driven by these physical processes. Forinstance, collating multiple image looks of a round object uncovers rotational symmetries in anappropriate feature space derived from the images. The use of topological invariants allows oneto infer that the object is round by reasoning about its feature space. The fact that sonar targetsignatures are (mostly) translation invariant in range can also be deduced from topologicalinvariants. Our approach is therefore both data-driven and mathematically principled ??? thisallows it to look for structure and symmetry within the data, which will be related to complexphysical features. As an additional benefit, topological approaches can be used to identifyweaknesses and limitations in simulated datasets.

Document Details

Document Type
DoD Grant Award
Publication Date
Jul 26, 2018
Source ID
N000141812541

Entities

People

  • Michael Robinson

Organizations

  • American University
  • Office of Naval Research
  • United States Navy

Tags

Readers

  • Data Mining and Knowledge Discovery.
  • Graph Algorithms and Convex Optimization.
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • Space
  • Space - Space Objects