UC Riverside

As the use of machine and deep learning becomes increasingly ubiquitous, there is also a need to address the most common challenges in real-world applications of deep learning. This thesis presents a study of four applications, for building solutions with informed design choices addressing major challenges such as limited and diverse data problems, tackling artifacts, and data imbalance - to name a few.

The first application deals with developing generative models to reproduce and characterize the visual manifestation of unapparent radiogenomic features in MR imaging for a rare mutation in Glioblastoma patients. This research reports three contributions in this study: (1) a novel generative approach that models tumor macro-features in this limited, diverse, and unbalanced dataset, (2) a novel approach to computationally model tumor invasion properties, and (3) a new metric to evaluate GANs.

The second application has an astrobiological utility to study one of the most definitive biogenic signs of early life on Earth, the double-rippled bedforms (DRBs) from the Ediacaran Member (550–555 Ma) of South Australia. We develop automated tools for this analysis while tackling the challenges of an uncurated, limited, and artifact-ridden dataset. We develop a new automated tool for astrobiological investigation that robustly detects miniature biosignatures while tackling these challenges.

The third application addresses the challenge of video-based human recognition using a limited feature set under extreme imaging distortions, imaging range, lack of frames, arbitrary pose, occlusions, air turbulence, and changing clothes. We propose two novel body-based biometrics, one using 2D binary silhouette, and one using 2D RGB images, that can tackle the aforementioned challenges while using minimal information at a time. Extensive evaluation reveals the efficacy of each of these biometrics under varying imaging conditions.

The fourth application is to develop novel generative methods to synthesize full-atom 3D protein conformations for the Aβ42 monomer. An intrinsically disordered protein, Aβ42 has a highly dynamic nature and a complex feature space. We propose a novel sampling technique to synthesize novel conformations that can capture the high variability in the protein features and has shown to be highly effective in sampling stable, diverse, unique, and biologically useful conformations under varying force fields.