Information Geometry and Information Structure
Abstract
This proposed effort will investigate how to extract and fully utilize physics-based features in acoustic data in order to improve t,he signal processing algorithms and representations of objects. This effort will be guided by the principles of information geometry, of objects in representations and the information structure in data-driven learning models. Current signal processing algorithms co,ncentrate raw acoustic energy of target responses into a small number of ?pixels? that form a recognizable shape. Using learning alg,orithms, this shape feature can be further sparsified with higher level of abstraction. We hypothesize that selective concentration,of relevant energy increases the local signal to noise ratio (SNR), and that improved SNR allows for robust feature extraction acros,s diverse data collection environments. We will employ a three-pronged approach for this effort. They are 1) build explainable learn,ing models by designing attribute-specific data collection experiments, 2) use data-driven approaches for further sparsification of,high-level features, and 3) design signal processing algorithms via modular response analysis of trained models.ARL/PSU has develope,d, with internal funds, in-air synthetic aperture sonar data collection framework that enables precisely controlled experiments. Thi,s framework provides a repeatable and flexible means for collecting experimental data to investigate information structure free from, the typical challenges of underwater data collection. This data will be used with learning algorithms to train for specific parts o,f the overall response. Sufficiently generalized component responses learned from basic objects in this controlled manner will be mo,re explainable. We will use machine learning algorithms to capture the feature extraction part of the trained neural network for pur,poses of investigating the information structure. Sparse autoencoders, for example, are learning algorit,information structure of the unlabeled input data into sparse codes. By encouraging the code nodes to be sparse, the network learns,to represents the input data in a compressed form. Using this property as a tool, we will provide the network with increasingly abst,racted representations learned by previous stage of signal processing and feature extraction, starting from the representation stage, such as image or acoustic color. Instead of collecting a large quantity of data with random configuration variations, which will be, labor-intensive to label, we will use small datasets with data augmentation techniques such as random partial occlusion to encourag,e a small network to learn as much about the data as possible. Visualization techniques such as saliency maps will be used to valida,te that the learned network is capturing the expected signatures of the input data. Small networks with fewer layers and channels ar,e also more practical for analysis and suitable for building explainable models [D?souza 2020].The third component is designing phys,ics-based signal processing algorithms by analyzing the information structure learned by the model. We will use learning algorithms,as a tool to improve our understanding of the information structure that are obvious in the representation and not captured by engin,eered features, but still extractable by learning algorithms. Emphasis will be put on identifying generalizable physical principles,that are captured by the trained network, so that it will be applicable to more general cases outside of the training data set. DIST,RIBUTION STATEMENT A. APPROVED FOR PUBLIC RELEASE. DISTRIBUTION IS UNLIMITED.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 05, 2022
- Source ID
- N000142212627
Entities
People
- Joonho Park
Organizations
- Office of Naval Research
- Pennsylvania State University
- United States Navy