Bita Rouhani
University of California, San Diego
bita@ucsd.edu
Bio
Bita Rouhani is a last year Ph.D. candidate at University of California San Diego with an expected degree in Electrical and Computer Engineering. Bita has received her master degree in Computer Engineering from Rice University. She is a recipient of Microsoft Ph.D. Fellowship. Her research interests include reconfigurable computing distributed optimization big data analytics and deep learning.
Succinct and Assured Machine Learning: Training and Execution
Succinct and Assured Machine Learning: Training and Execution
Physical viability and safety consideration are key factors in devising machine learning systems that are both sustainable and trustworthy. Learning and analysis of massive data is a trend that is ubiquitous among different computing platforms ranging from smartphones and Internet-of-Thing (IoT) devices to personal computers and many-core cloud servers. Concerns over the functionality (accuracy), physical performance, and reliability (safety) are major challenges in building automated learning systems that can be employed in real-world settings. My research work aims to address these three critical aspects of emerging computing scenarios: the functionality, physical performance, and reliability. What makes my work distinctive is that it provides holistic automated solutions that simultaneously capture the best of Computer Architecture, Machine Learning (ML), and Security fields. I introduced, realized, and automated a resource-aware deep learning framework (called Deep3) that enables succinct training and execution of DL models by simultaneously leveraging three levels of parallelism: data, network, and hardware. The resource efficiency of my solution for the first time enables dynamic training and execution of DL networks in resource constraints settings applied to both distributed cloud servers with limited communication bandwidth as well as embedded devices with bounded computational power (e.g., smartphones). I further enriched my proposed DL framework by providing algorithmic and custom hardware-accelerated tools to ensure the reliability of model prediction in the face of adversarial attacks. The reliability aspect of my developed solution in turn empowers the use of assured DL models in sensitive scenarios where data privacy and/or robustness against adversarial samples is crucial.