Projected Belief Networks for Classification of Underwater Acoustic Signals
Abstract
Title: Projected Belief Networks for Classifying Ocean Acoustic EventsThis pr made at Fraunhofer FKIE, called "Projected Belief Network", in order to improve the state of the art in classifying underwater acoustic events. Neural networks can be classified into two broad groups, discriminative networks and generative networks. State of the art (SOA) deep neural network classifiers are of the discriminative variety, whereas generative networks have been primarily usedrmance. Discriminative deep neural networks have been brilliantly successful, but they owe their success to the availability of very large training data sets and have no inherent model of the data generation process. As a result of this shortcoming, discriminative networks fail to deliver promised performance on smaller military-typedata sets and can be easily fooled, for example by the method of adversarial sampling. Generative networks, on the other hand are based on a model of the underlying data so are difficultto fool, but fail to deliver state of the art classification performance due to the inherent difficulty of generative modeling. What is needed is a generative network that has excellent classification performance. The projected belief network (PBN), developed atFraunhofer FKIE, promises to merge discriminative and generative networks into a single network. Since the PBN is the only type ofgenerative network based on a feed-forward network (as are discriminative neural networks), it is can share an embodiment with a discriminative network. Theoretically, the PBN can be trained to achieve state-of-the-art classification performance, yet can at the same time possess a model of the underlying data. To generate synthetic data or re-generate existing visible data, the PBN employs tiables at any layer. The quality of reconstructed data testifies to the descriptive information present in any layer. Experiments have shown that the presence of descriptive information does not disturb the ability of the network to classify, but in fact may improve the generalization capability. In addition to classifying, a trained PBN can be used to generate synthetic data and to reject out-of-set samples, something that discriminative networks are not suited to do.This project seeks to further develop the PBN and apply it to underwater acoustic data. In an initial project phase, a consultation with scientists at Naval Undersea Warfare Center would seek to direct the research toward data of interest to US Navy. The bulk of the project effort will be directed toward optimizingPBN performance. Although a PBN is based on a conventional network, it has fundamentally different properties. As such, it will take extensive experimentation to find the network structure, complexity, and regularization methods that optimize the PBN performance and strike the best trade-offs between discriminative and generative performance. The networks will be evaluated according to the classification performance, the ability to generate realistic synthetic data, computational efficiency, and generalization capability. Optimization of software for graphical processing units (GPUs) may also be necessary. Rather than using large open data sets, the experiments will focus on small difficult-to obtain data sets, such as seen in military applications. Additionally, some open theoretical questions will be addressed. For example, the PBN may solve an important problem in information theory the maximization of mutual information between the network input and output. Theoretical and experimental results will be submitted to highly-regarded journals.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2021
- Source ID
- N629092112024
Entities
People
- Paul Baggenstoss
Organizations
- Fraunhofer Society
- Office of Naval Research
- United States Navy