Learned Approach to Sonar Signal Interpretation and Evaluation (LASSIE)

Charles River Analytics proposes to design and demonstrate a Learned Approach to Sonar Signal Interpretation and Evaluation (LASSIE), a research program to compare the effectiveness of different types of deep learning (DL) in solving the problem of clutter reduction in MIMO sonar and to design an improved signal processing chain which is uniquely adapted for deep learning. LASSIE employs advanced probabilistic models to featurize and train deep neural networks. By combining a recurrent neural network with deep generative models, it is possible to discover representative probability distributions for non-target objects and thereby reject input which is non-conforming. The use of a generative model also has the advantage that novel environmental conditions can be discovered and trained without the need for experimental data collected in the field. Our approach includes echo synchronization and geolocation as well as a combination of unsupervised and supervised learning strategies for target detection.