Thomas Powers Research Assistant & Research Assistant tcpowers@apl.washington.edu |

Publications |
2000-present and while at APL-UW |

Deep generative acoustic models for real-time sonar systems Powers, T., and D.W. Krout, "Deep generative acoustic models for real-time sonar systems," Proc., 21st International Conference on Information Fusion (FUSION), 10-13 July, Cambridge, UK, doi:10.23919/ICIF.2018.8455375 (IEEE, 2018). |
More Info |
10 Jul 2018 |
|||||||

High-fidelity acoustic models are crucial to the performance of sonar systems since they identify where there is signal excess in the surrounding environment. Sonar operators use these models to optimize sonar parameters in applications like target detection and tracking. Unfortunately, high-fidelity generative models like Comprehensive Acoustic System Simulation (CASS) are very computationally intensive and can not be run in real time. One way around this limitation is to pre-compute maps for a region given expected environmental parameters, but this approach is impractical in highly variable littoral regions. Instead, we propose to emulate a high-fidelity acoustic model with a deep neural network (DNN). DNNs are very efficient at test time, requiring only a few iterations of simple linear and nonlinear operations, and can easily be run in real-time on a sonar platform. Deep models continue to advance the state-of-the-art in fields like speech and audio processing, computer vision, and natural language processing. These advances motivate the use of deep learning for a real-time generative acoustic model. We train a deep feedforward sigmoid network on a dataset generated by CASS and demonstrate promising results. |

Constrained robust submodular sensor selection with application to multi static sonar arrays Powers, T., D.W. Trout, J. Blimes, and L. Atlas, "Constrained robust submodular sensor selection with application to multi static sonar arrays," IET Radar Sonar Navig., 11, 1776-1781, doi:10.1049/iet-rsn.2017.0075, 2017. |
More Info |
1 Dec 2017 |
|||||||

The authors develop a framework to select a subset of sensors from a field in which the sensors have an ingrained independence structure. Given an arbitrary independence pattern, the authors construct a graph that denotes pairwise independence between sensors, which means those sensors may operate simultaneously without interfering. The set of all fully-connected subgraphs (cliques) of this independence graph forms the independent sets of matroids over which the authors maximise the average and minimum of a set of submodular objective functions. The average case is submodular, so it can be approximated. The minimum case is both non-submodular and inapproximable. The authors propose a novel algorithm GENSAT that exploits submodularity and, as a result, returns a near-optimal solution with approximation guarantees on a relaxed problem that are within a small factor of the average case scenario. The authors apply this framework to ping sequence optimisation for active multistatic sonar arrays by maximising sensor coverage for average and minimum case scenarios and derive lower bounds for minimum probability of detection for a fractional number of targets. In these ping sequence optimisation simulations, GENSAT exceeds the fractional lower bounds and reaches near-optimal performance, and submodular function optimisation vastly outperforms traditional approaches and nearly achieves optimal performance. |

Building recurrent networks by unfolding iterative thresholding for equential sparse recovery Wisdom, S., T. Powers, J. Pitton, and L. Atlas, "Building recurrent networks by unfolding iterative thresholding for equential sparse recovery," Proc., IEEE International Conference on Acoustics, Speech and Signal Processing, 5-9 March, New Orleans, LA, 4346-4350, doi:10.1109/ICASSP.2017.7952977 (IEEE, 2017). |
More Info |
19 Jun 2017 |
|||||||

Historically, sparse methods and neural networks, particularly modern deep learning methods, have been relatively disparate areas. Sparse methods are typically used for signal enhancement, compression, and recovery, usually in an unsupervised framework, while neural networks commonly rely on a supervised training set. In this paper, we use the specific problem of sequential sparse recovery, which models a sequence of observations over time using a sequence of sparse coefficients, to show how algorithms for sparse modeling can be combined with supervised deep learning to improve sparse recovery. Specifically, we show that the iterative soft-thresholding algorithm (ISTA) for sequential sparse recovery corresponds to a stacked recurrent neural network (RNN) under specific architecture and parameter constraints. Then we demonstrate the benefit of training this RNN with backpropagation using supervised data for the task of column-wise compressive sensing of images. This training corresponds to adaptation of the original iterative thresholding algorithm and its parameters. Thus, we show by example that sparse modeling can provide a rich source of principled and structured deep network architectures that can be trained to improve performance on specific tasks. |