Exploitation of Ephemeral Features in Sonar Classification Algorithms

There is a current U.S. Navy need for robust target classification using both passive and active sonar.  The importance of this problem has produced a rich history of classification strategies and contact features, such as perceptual-based features from timbre content and time-frequency analysis, and subband attack time statistics of a filterbank analysis.  However, in challenging operating environments a target may only exist for fleeting moments, limiting the information collected by the sonar receiver.  Further, man-made objects of interest to the U.S. Navy, such as submarines and torpedoes, may take intentionally deceptive and evasive measures to limit detection.  These scenarios complicate the classification process because not all potential features are available for all sonar contacts.  Instead, some features may be frequently available, while others are rarely available, i.e. ephemeral.  Although recent machine learning methods, such as deep neural networks, have demonstrated state-of-the-art classification performance, it is well-known that missing features at test time represent a “nightmare scenario” capable of completely distorting classification outputs. While simple data imputation methods, such as substituting missing data with its training-data mean (or worse, with zeros) may seem like a reasonable first approach, it is unsurprising that such synthetic features work poorly when used with highly tuned machine learning algorithms that have never been presented with such feature combinations during training. In the proposed work, we will develop and evaluate automated classification techniques that mitigate the deleterious effects of missing features. Further, because missing features do not necessarily occur uniformly at random, the availability or absence of a feature may itself be an informative event that can be exploited for improved classification performance. The primary objective of the Phase 1 effort is to develop a sophisticated sonar classification strategy that is robust to ephemeral features.  In the proposed work, we will augment off-the-shelf neural network classification methods with robustness to missing features and the ability to optionally exploit the non-uniformity of a missing data process for improved classification performance. We will consider a fully general data availability model capable of supporting features that are always available, sometimes available, and rarely available (i.e., ephemeral features) as special cases. In this effort, our approach to the topic of the ephemeral feature will be divided into three stages of increasing sophistication: Stage-1: Improved data imputation methods Stage-2: End-to-end classification strategies robust to missing data Stage-3: Exploitation of the missing data process