Improving Deep Representations by Incorporating Domain Knowledge and Modularization for Synthetic Aperture Radar and Physiological Data

Abstract

Machine Learning (ML) using Artificial Neural Networks (ANNs), referred to as Deep Learning (DL), is a very popular and powerful method of statistical inference. A primary advantage of deep-learning has been the automatic learning of informa- tive features (that encodes the data referred to as deep-representations henceforth) based on gradient-descent optimization of an objective function. While DL is ap- plicable to problem domains where hand-crafted features are not readily available, its performance is critically dependent on other factors like dataset size and model architecture. Despite recent advances in the field, the question of how to modify the DL framework to incorporate domain knowledge or to disentangle factors of varia- tion warrants more research. Until recently, most popular works in the DL literature have primarily employed inductive-bias of architectures (e.g., translational invariance in convolutional neural-nets) and relied on the availability of large labeled datasets for improved representation learning. Unfortunately, curating such large datasets is costly and not practical for many application areas. In this dissertation, we study methods to improve learned representations by incorporating domain knowledge into the learning process and through disentangling factors of variation.

First, we present a sparse-modeling based data augmentation method for tomo- graphic images and use it to incorporate domain knowledge of Synthetic Aperture

Radar (SAR) target phenomenology into deep representations. We validate the im- provements in learned representations by using them for a benchmark classification problem of Automatic Target Recognition (ATR) where we establish new state-of-the- art on subsampled datasets. Second, we propose a DL-based hierarchical modeling strategy for a physiological signal generation process which in turn can be used for data augmentation. Based on the physiology of cardiovascular system function, we propose a modularized hierarchical generative model and then impose explicit regular- izing constraints on each module using multi-objective loss functions. This generative model, called CardioGen, is evaluated by its ability to augment real data while train- ing DL based models. The proposed approach showed performance improvements.

Third, we propose a hierarchical deep-generative model for SAR imagery that jointly captures the underlying structure of multiple resolutions of SAR images. We utilize this model, called MrSARP, to super-resolve lower resolution magnitude images to a higher resolution. We evaluate the model’s performance using the three standard error metrics used for evaluating super-resolution performance on simulated data. Fourth, we propose a framework for learning a sufficient statistic of the data for a given downstream inference task. We design and train a DL model that encodes the Photoplethysmography (PPG) signal to a sufficient statistic and decodes it back to a task-specific PPG-like signal assuming it will be used for a fixed RR-tachogram prediction task. Compression and privacy-preservation can be a useful side-benefit of having such a downstream task. The learned deep representations of PPG data are validated using tachogram prediction error as well as its performance on the sub-task of stress estimation.