Mrs. Luisa Polania Cabrera
Institution: DePaul University
Department: School of Computing and Digital Media
Proposed research ideas:
I propose three research ideas: – Development of deep learning architectures for multimodal learning problems: Real-world scenarios often involve acquisition of data from multiple sources. However, fusing information from different sources is a difficult task due to differences in the statistical properties of the modalities. I propose to develop deep learning models that simultaneously learn temporal dependencies and a joint representation between modalities. Prior work in this area include one of my previous projects at the Palo Alto Research Center related to activity recognition in the process of insulin self-injection using both accelerometer data and video as sources of information. – Exploiting deep learning architectures and restricted Boltzmann machines in compressed sensing: One of my research areas is related to exploiting signal and measurement structure in compressed sensing. Some Lawrence Berkeley National Laboratory (LBNL) researchers (Alexander Pines, Vikram Bajaj, Grant T. Gullberg, Chao Yang) have lead research on the application of compressed sensing to medical imaging. Even though I have not worked on medical imaging applications before, I believe the application of my algorithms to this domain will lead to significant improvements in reconstruction performance. My algorithms exploit the representational power of restricted Boltzmann machines and deep learning architectures to model the prior distribution of the sparsity pattern of signals and to leverage statistical dependencies between elements of the signal support. – Development of deep learning ranking models: Even though ordinal ranking problems are an important research area with multiple applications in different domains, there are limited existing deep learning models for ranking. The goal of an ordinal ranking problem is to classify patterns using a categorical scale which shows a natural order between labels, but not a meaningful numeric difference between them. These kind of problems are not successfully addressed by the standard regression and classification deep learning architectures since the absolute difference of output values is nearly meaningless and only their relative order matters. I propose to develop novel deep learning ranking models.
My main motivation is to establish research collaborations with world-class researchers from the Lawrence Berkeley National Laboratory (LBNL) and to work on challenging problems that arise in different scientific and engineering fields, such as material science, biology, climate modeling, and many others. So far, I have only worked on applications related to healthcare, wellness, and the insurance industry (e.g. prediction of roof condition from property survey images). This program will give me and my future students the opportunity to increase the impact of our research by applying our technical skills to a broader range of applications. The available data sources and computational resources at LBNL are additional motivation factors to participate in this program.